From 1a6c021cd102b48b44e006decebc165062df8a95 Mon Sep 17 00:00:00 2001
From: 34j <55338215+34j@users.noreply.github.com>
Date: Mon, 3 Apr 2023 12:22:51 +0900
Subject: [PATCH] feat: add quickvc, fix usage of ContentVec, remove onnx
 support (#197)

BREAKING CHANGE: The module structure has been completely changed, and some of the default values for configurations have also been changed.
---
 .idea/workspace.xml                           |  89 ++-
 LICENSE                                       | 201 ++++++
 README.md                                     |  65 +-
 src/so_vits_svc_fork/__main__.py              |  60 +-
 src/so_vits_svc_fork/data_utils.py            |   4 +-
 src/so_vits_svc_fork/f0.py                    | 237 ++++++++
 src/so_vits_svc_fork/gui.py                   |  21 +-
 src/so_vits_svc_fork/hparams.py               |  38 ++
 src/so_vits_svc_fork/hubert/hubert_model.py   | 223 -------
 .../hubert/hubert_model_onnx.py               | 219 -------
 .../hubert/put_hubert_ckpt_here               |   0
 .../inference/{infer_tool.py => core.py}      |  88 +--
 .../{inference_main.py => inference/main.py}  |   2 +-
 src/so_vits_svc_fork/models.py                | 519 ----------------
 .../{hubert => modules/decoders}/__init__.py  |   0
 src/so_vits_svc_fork/modules/decoders/f0.py   |  45 ++
 .../modules/decoders/hifigan/__init__.py      |   3 +
 .../decoders/hifigan/_models.py}              | 361 +----------
 .../modules/decoders/hifigan/_utils.py        |  15 +
 .../modules/decoders/mb_istft/__init__.py     |  15 +
 .../modules/decoders/mb_istft/_generators.py  | 376 ++++++++++++
 .../modules/decoders/mb_istft/_loss.py        |  11 +
 .../modules/decoders/mb_istft/_pqmf.py        | 130 ++++
 .../modules/decoders/mb_istft/_stft.py        | 248 ++++++++
 .../modules/decoders/mb_istft/_stft_loss.py   | 142 +++++
 .../modules/descriminators.py                 | 177 ++++++
 src/so_vits_svc_fork/modules/encoders.py      | 136 +++++
 src/so_vits_svc_fork/modules/flows.py         |  48 ++
 .../modules/mel_processing.py                 |   4 +-
 src/so_vits_svc_fork/modules/synthesizers.py  | 233 +++++++
 src/so_vits_svc_fork/onnx_export.py           |  65 --
 src/so_vits_svc_fork/onnxexport/model_onnx.py | 404 ------------
 .../{onnxexport => preprocessing}/__init__.py |   0
 .../config_templates/quickvc.json             |  69 +++
 .../so-vits-svc-4.0v1-legacy.json}            |   0
 .../config_templates/so-vits-svc-4.0v1.json   |  62 ++
 .../preprocess_flist_config.py                |   6 +-
 .../preprocess_hubert_f0.py                   |  63 +-
 .../preprocess_resample.py                    |   0
 .../preprocess_speaker_diarization.py         |   0
 .../{ => preprocessing}/preprocess_split.py   |   0
 .../{ => preprocessing}/preprocess_utils.py   |   0
 src/so_vits_svc_fork/spec_gen.py              |  23 -
 src/so_vits_svc_fork/train.py                 | 157 +++--
 src/so_vits_svc_fork/utils.py                 | 575 ++++++------------
 src/so_vits_svc_fork/vdecoder/__init__.py     |   0
 .../vdecoder/hifigan/__init__.py              |   0
 src/so_vits_svc_fork/vdecoder/hifigan/env.py  |  15 -
 .../vdecoder/hifigan/nvSTFT.py                | 162 -----
 .../vdecoder/hifigan/utils.py                 |  71 ---
 tests/test_main.py                            |  30 +-
 51 files changed, 2736 insertions(+), 2676 deletions(-)
 create mode 100644 src/so_vits_svc_fork/f0.py
 create mode 100644 src/so_vits_svc_fork/hparams.py
 delete mode 100644 src/so_vits_svc_fork/hubert/hubert_model.py
 delete mode 100644 src/so_vits_svc_fork/hubert/hubert_model_onnx.py
 delete mode 100644 src/so_vits_svc_fork/hubert/put_hubert_ckpt_here
 rename src/so_vits_svc_fork/inference/{infer_tool.py => core.py} (91%)
 rename src/so_vits_svc_fork/{inference_main.py => inference/main.py} (99%)
 delete mode 100644 src/so_vits_svc_fork/models.py
 rename src/so_vits_svc_fork/{hubert => modules/decoders}/__init__.py (100%)
 create mode 100644 src/so_vits_svc_fork/modules/decoders/f0.py
 create mode 100644 src/so_vits_svc_fork/modules/decoders/hifigan/__init__.py
 rename src/so_vits_svc_fork/{vdecoder/hifigan/models.py => modules/decoders/hifigan/_models.py} (50%)
 create mode 100644 src/so_vits_svc_fork/modules/decoders/hifigan/_utils.py
 create mode 100644 src/so_vits_svc_fork/modules/decoders/mb_istft/__init__.py
 create mode 100644 src/so_vits_svc_fork/modules/decoders/mb_istft/_generators.py
 create mode 100644 src/so_vits_svc_fork/modules/decoders/mb_istft/_loss.py
 create mode 100644 src/so_vits_svc_fork/modules/decoders/mb_istft/_pqmf.py
 create mode 100644 src/so_vits_svc_fork/modules/decoders/mb_istft/_stft.py
 create mode 100644 src/so_vits_svc_fork/modules/decoders/mb_istft/_stft_loss.py
 create mode 100644 src/so_vits_svc_fork/modules/descriminators.py
 create mode 100644 src/so_vits_svc_fork/modules/encoders.py
 create mode 100644 src/so_vits_svc_fork/modules/flows.py
 create mode 100644 src/so_vits_svc_fork/modules/synthesizers.py
 delete mode 100644 src/so_vits_svc_fork/onnx_export.py
 delete mode 100644 src/so_vits_svc_fork/onnxexport/model_onnx.py
 rename src/so_vits_svc_fork/{onnxexport => preprocessing}/__init__.py (100%)
 create mode 100644 src/so_vits_svc_fork/preprocessing/config_templates/quickvc.json
 rename src/so_vits_svc_fork/{configs_template/config_template.json => preprocessing/config_templates/so-vits-svc-4.0v1-legacy.json} (100%)
 create mode 100644 src/so_vits_svc_fork/preprocessing/config_templates/so-vits-svc-4.0v1.json
 rename src/so_vits_svc_fork/{ => preprocessing}/preprocess_flist_config.py (92%)
 rename src/so_vits_svc_fork/{ => preprocessing}/preprocess_hubert_f0.py (60%)
 rename src/so_vits_svc_fork/{ => preprocessing}/preprocess_resample.py (100%)
 rename src/so_vits_svc_fork/{ => preprocessing}/preprocess_speaker_diarization.py (100%)
 rename src/so_vits_svc_fork/{ => preprocessing}/preprocess_split.py (100%)
 rename src/so_vits_svc_fork/{ => preprocessing}/preprocess_utils.py (100%)
 delete mode 100644 src/so_vits_svc_fork/spec_gen.py
 delete mode 100644 src/so_vits_svc_fork/vdecoder/__init__.py
 delete mode 100644 src/so_vits_svc_fork/vdecoder/hifigan/__init__.py
 delete mode 100644 src/so_vits_svc_fork/vdecoder/hifigan/env.py
 delete mode 100644 src/so_vits_svc_fork/vdecoder/hifigan/nvSTFT.py
 delete mode 100644 src/so_vits_svc_fork/vdecoder/hifigan/utils.py

diff --git a/.idea/workspace.xml b/.idea/workspace.xml
index 7b269da5..73a840f7 100644
--- a/.idea/workspace.xml
+++ b/.idea/workspace.xml
@@ -2,62 +2,36 @@
 <project version="4">
   <component name="ChangeListManager">
     <list default="true" id="7aacf003-229b-4eb0-80a4-ff105dc3c3d4" name="変更" comment="">
-      <change beforePath="$PROJECT_DIR$/.github/ISSUE_TEMPLATE/1-bug_report.md" beforeDir="false" afterPath="$PROJECT_DIR$/.github/ISSUE_TEMPLATE/1-bug_report.md" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.github/ISSUE_TEMPLATE/2-feature-request.md" beforeDir="false" afterPath="$PROJECT_DIR$/.github/ISSUE_TEMPLATE/2-feature-request.md" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.github/workflows/ci.yml" beforeDir="false" afterPath="$PROJECT_DIR$/.github/workflows/ci.yml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.github/workflows/hacktoberfest.yml" beforeDir="false" afterPath="$PROJECT_DIR$/.github/workflows/hacktoberfest.yml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.github/workflows/issue-manager.yml" beforeDir="false" afterPath="$PROJECT_DIR$/.github/workflows/issue-manager.yml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.github/workflows/labels.yml" beforeDir="false" afterPath="$PROJECT_DIR$/.github/workflows/labels.yml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.gitpod.yml" beforeDir="false" afterPath="$PROJECT_DIR$/.gitpod.yml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.idea/inspectionProfiles/Project_Default.xml" beforeDir="false" afterPath="$PROJECT_DIR$/.idea/inspectionProfiles/Project_Default.xml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.idea/inspectionProfiles/profiles_settings.xml" beforeDir="false" afterPath="$PROJECT_DIR$/.idea/inspectionProfiles/profiles_settings.xml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.idea/misc.xml" beforeDir="false" afterPath="$PROJECT_DIR$/.idea/misc.xml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.idea/modules.xml" beforeDir="false" afterPath="$PROJECT_DIR$/.idea/modules.xml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.idea/so-vits-svc-fork.iml" beforeDir="false" afterPath="$PROJECT_DIR$/.idea/so-vits-svc-fork.iml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.idea/vcs.xml" beforeDir="false" afterPath="$PROJECT_DIR$/.idea/vcs.xml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.idea/watcherTasks.xml" beforeDir="false" afterPath="$PROJECT_DIR$/.idea/watcherTasks.xml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.idea/workspace.xml" beforeDir="false" afterPath="$PROJECT_DIR$/.idea/workspace.xml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.pre-commit-config.yaml" beforeDir="false" afterPath="$PROJECT_DIR$/.pre-commit-config.yaml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/.readthedocs.yml" beforeDir="false" afterPath="$PROJECT_DIR$/.readthedocs.yml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/README_zh_CN.md" beforeDir="false" afterPath="$PROJECT_DIR$/README_zh_CN.md" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/commitlint.config.js" beforeDir="false" afterPath="$PROJECT_DIR$/commitlint.config.js" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/pyproject.toml" beforeDir="false" afterPath="$PROJECT_DIR$/pyproject.toml" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/renovate.json" beforeDir="false" afterPath="$PROJECT_DIR$/renovate.json" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/app.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/app.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/cluster/__init__.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/cluster/__init__.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/cluster/train_cluster.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/cluster/train_cluster.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/configs_template/config_template.json" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/configs_template/config_template.json" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/data_utils.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/data_utils.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/hubert/hubert_model.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/hubert/hubert_model.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/hubert/hubert_model_onnx.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/hubert/hubert_model_onnx.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/inference/infer_tool.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/inference/infer_tool.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/inference/infer_tool_grad.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/inference/infer_tool_grad.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/inference/slicer.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/inference/slicer.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/inference_main.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/inference_main.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/models.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/models.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/attentions.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/attentions.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/commons.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/commons.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/losses.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/losses.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/mel_processing.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/mel_processing.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/modules.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/modules.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/onnx_export.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/onnx_export.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/onnxexport/model_onnx.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/onnxexport/model_onnx.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/preprocess_flist_config.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/preprocess_flist_config.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/preprocess_hubert_f0.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/preprocess_hubert_f0.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/resample.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/resample.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/spec_gen.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/spec_gen.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/train.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/train.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/utils.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/utils.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/vdecoder/hifigan/env.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/vdecoder/hifigan/env.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/vdecoder/hifigan/models.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/vdecoder/hifigan/models.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/vdecoder/hifigan/nvSTFT.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/vdecoder/hifigan/nvSTFT.py" afterDir="false" />
-      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/vdecoder/hifigan/utils.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/vdecoder/hifigan/utils.py" afterDir="false" />
+      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/__main__.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/__main__.py" afterDir="false" />
+      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/gui.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/gui.py" afterDir="false" />
+      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/onnx/__init__.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/onnx/__init__.py" afterDir="false" />
+      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/onnx/model_onnx.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/onnx/_model.py" afterDir="false" />
+      <change beforePath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/onnx/onnx_export.py" beforeDir="false" afterPath="$PROJECT_DIR$/src/so_vits_svc_fork/modules/onnx/_export.py" afterDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/1.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/10.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/2.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/3.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/4.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/5.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/6.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/7.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/8.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/9.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/dataset_raw/34j/nested/あ.wav" beforeDir="false" />
+      <change beforePath="$PROJECT_DIR$/tests/test_main.py" beforeDir="false" afterPath="$PROJECT_DIR$/tests/test_main.py" afterDir="false" />
     </list>
     <option name="SHOW_DIALOG" value="false" />
     <option name="HIGHLIGHT_CONFLICTS" value="true" />
     <option name="HIGHLIGHT_NON_ACTIVE_CHANGELIST" value="false" />
     <option name="LAST_RESOLUTION" value="IGNORE" />
   </component>
+  <component name="FileTemplateManagerImpl">
+    <option name="RECENT_TEMPLATES">
+      <list>
+        <option value="Python Script" />
+      </list>
+    </option>
+  </component>
   <component name="Git.Settings">
     <option name="RECENT_GIT_ROOT_PATH" value="$PROJECT_DIR$" />
   </component>
@@ -65,6 +39,9 @@
     <option name="stateVersion" value="1" />
   </component>
   <component name="ProjectId" id="2N3U7T2ZqSld9sk8NQi5nSwUOO9" />
+  <component name="ProjectLevelVcsManager">
+    <ConfirmationsSetting value="2" id="Add" />
+  </component>
   <component name="ProjectViewState">
     <option name="hideEmptyMiddlePackages" value="true" />
     <option name="showLibraryContents" value="true" />
@@ -79,6 +56,15 @@
     &quot;nodejs_package_manager_path&quot;: &quot;npm&quot;
   }
 }</component>
+  <component name="RecentsManager">
+    <key name="MoveFile.RECENT_KEYS">
+      <recent name="C:\Users\m\MyMain\dev\so-vits-svc-fork\src\so_vits_svc_fork\modules" />
+      <recent name="C:\Users\m\MyMain\dev\so-vits-svc-fork\src\so_vits_svc_fork\inference" />
+      <recent name="C:\Users\m\MyMain\dev\so-vits-svc-fork\src\so_vits_svc_fork\onnxexport" />
+      <recent name="C:\Users\m\MyMain\dev\so-vits-svc-fork\src\so_vits_svc_fork\preprocessing" />
+      <recent name="C:\Users\m\MyMain\dev\so-vits-svc-fork\src\so_vits_svc_fork\modules\decoders" />
+    </key>
+  </component>
   <component name="RunManager" selected="Python.so-vits-svc-fork">
     <configuration name="so-vits-svc-fork" type="PythonConfigurationType" factoryName="Python">
       <module name="so-vits-svc-fork" />
@@ -140,6 +126,9 @@
       <updated>1678892092249</updated>
       <workItem from="1678892093553" duration="810000" />
       <workItem from="1678932243084" duration="593000" />
+      <workItem from="1680174456649" duration="1005000" />
+      <workItem from="1680251014707" duration="2800000" />
+      <workItem from="1680319074742" duration="3292000" />
     </task>
     <servers />
   </component>
diff --git a/LICENSE b/LICENSE
index 764c22e5..0bcb3ed1 100644
--- a/LICENSE
+++ b/LICENSE
@@ -20,3 +20,204 @@ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 SOFTWARE.
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
diff --git a/README.md b/README.md
index 80750b39..8d2cc828 100644
--- a/README.md
+++ b/README.md
@@ -37,13 +37,15 @@ A fork of [`so-vits-svc`](https://github.com/svc-develop-team/so-vits-svc) with
 ## Features not available in the original repo
 
 - **Realtime voice conversion** (enhanced in v1.1.0)
-- More accurate pitch estimation using CREPE
-- GUI available
-- Unified command-line interface (no need to run Python scripts)
+- Integrates [`QuickVC`](https://github.com/quickvc/QuickVC-VoiceConversion)
+- Fixed misuse of `ContentVec` in the original repository.[^c]
+- More accurate pitch estimation using [`CREPE`](https://github.com/marl/crepe/).
+- GUI and unified CLI available
 - Ready to use just by installing with `pip`.
-- Automatically download pretrained base model and HuBERT model
+- Automatically download pretrained models.
 - Code completely formatted with black, isort, autoflake etc.
-- Other minor differences
+
+[^c]: [#206](https://github.com/34j/so-vits-svc-fork/issues/206)
 
 ## Installation
 
@@ -53,23 +55,53 @@ A fork of [`so-vits-svc`](https://github.com/svc-develop-team/so-vits-svc) with
   <img src="https://img.shields.io/badge/.bat-download-blue?style=flat-square&logo=windows" alt="Download .bat">
 </a>
 
-### [Creating a Virtual Environment](https://github.com/34j/so-vits-svc-fork/wiki#creating-a-virtual-environment)
+### Manual installation
+
+<details>
+  <summary>Creating a virtual environment</summary>
+
+Windows:
+
+```shell
+py -3.10 -m venv venv
+venv\Scripts\activate
+```
+
+Linux/MacOS:
+
+```shell
+python3.10 -m venv venv
+source venv/bin/activate
+```
+
+Anaconda:
+
+```shell
+conda create -n so-vits-svc-fork python=3.10 pip
+conda activate so-vits-svc-fork
+```
 
-### Install
+Installing without creating a virtual environment may cause a `PermissionError` if Python is installed in Program Files, etc.
+
+</details>
 
 Install this via pip (or your favourite package manager that uses pip):
 
 ```shell
 python -m pip install -U pip setuptools wheel
-pip install -U torch torchaudio --index-url https://download.pytorch.org/whl/cu117
+pip install -U torch torchaudio --index-url https://download.pytorch.org/whl/cu118
 pip install -U so-vits-svc-fork
 ```
 
-- If no GPU is available, simply remove `pip install -U torch torchaudio --index-url https://download.pytorch.org/whl/cu117`.
-- If you are using an AMD GPU on Linux, replace `--index-url https://download.pytorch.org/whl/cu117` with `--index-url https://download.pytorch.org/whl/rocm5.4.2`. AMD GPUs are not supported on Windows (#120).
+<details>
+  <summary>Notes</summary>
+
+- If no GPU is available, simply remove `pip install -U torch torchaudio --index-url https://download.pytorch.org/whl/cu118`.
+- If you are using an AMD GPU on Linux, replace `--index-url https://download.pytorch.org/whl/cu118` with `--index-url https://download.pytorch.org/whl/rocm5.4.2`. AMD GPUs are not supported on Windows ([#120](https://github.com/34j/so-vits-svc-fork/issues/120)).
 - If `fairseq` raises an error:
   - If it prompts [`Microsoft C++ Build Tools`](https://visualstudio.microsoft.com/visual-cpp-build-tools/) is not installed. please install it.
   - If it prompts that some dll is missing, reinstalling `Microsoft Visual C++ 2022` and `Windows SDK` may help.
+  </details>
 
 ### Update
 
@@ -98,13 +130,13 @@ svcg
 - Realtime (from microphone)
 
 ```shell
-svc vc --model-path <model-path>
+svc vc
 ```
 
 - File
 
 ```shell
-svc --model-path <model-path> source.wav
+svc infer source.wav
 ```
 
 [Pretrained models](https://huggingface.co/models?search=so-vits-svc-4.0) are available on HuggingFace.
@@ -119,8 +151,8 @@ svc --model-path <model-path> source.wav
 #### Before training
 
 - If your dataset has BGM, please remove the BGM using software such as [Ultimate Vocal Remover](https://ultimatevocalremover.com/). `3_HP-Vocal-UVR.pth` or `UVR-MDX-NET Main` is recommended. [^1]
-- If your dataset is a long audio file with multiple speakers, use `svc sd` to split the dataset into multiple files (using `pyannote.audio`). Further manual classification may be necessary due to accuracy issues. If speakers speak with a variety of speech styles, set --min-speakers larger than the actual number of speakers. Due to unresolved dependencies, please install `pyannote.audio` manually: `pip install pyannote-audio`.
 - If your dataset is a long audio file with a single speaker, use `svc split` to split the dataset into multiple files (using `librosa`).
+- If your dataset is a long audio file with multiple speakers, use `svc sd` to split the dataset into multiple files (using `pyannote.audio`). Further manual classification may be necessary due to accuracy issues. If speakers speak with a variety of speech styles, set --min-speakers larger than the actual number of speakers. Due to unresolved dependencies, please install `pyannote.audio` manually: `pip install pyannote-audio`.
 
 [^1]: https://ytpmv.info/how-to-use-uvr/
 
@@ -130,7 +162,7 @@ svc --model-path <model-path> source.wav
 [![Open In Paperspace](https://img.shields.io/badge/Open%20in-Paperspace-blue?style=flat-square&logo=paperspace)](https://console.paperspace.com/github/34j/so-vits-svc-fork-paperspace/blob/main/so-vits-svc-fork-4.0-paperspace.ipynb)
 [![Paperspace Referral](<https://img.shields.io/badge/Referral%20($10)-9VJN74I-blue?style=flat-square&logo=paperspace>)](https://www.paperspace.com/?r=9VJN74I)[^p]
 
-If you do not have access to a GPU with more than 10 GB of VRAM, the free plan of Google Colab is recommended for light users and the Pro/Growth plan of Paperspace is recommended for heavy users. Conversely, if you have a high-end GPU, the use of cloud services is not recommended.
+If you do not have access to a GPU with more than 10 GB of VRAM, the free plan of Google Colab is recommended for light users and the Pro/Growth plan of Paperspace is recommended for heavy users. Conversely, if you have access to a high-end GPU, the use of cloud services is not recommended.
 
 [^p]: If you register a referral code and then add a payment method, you may save about $5 on your first month's monthly billing. Note that both referral rewards are Paperspace credits and not cash. It was a tough decision but inserted because debugging and training the initial model requires a large amount of computing power and the developer is a student.
 
@@ -148,8 +180,9 @@ svc train -t
 #### Notes
 
 - Dataset audio duration per file should be <~ 10s or VRAM will run out.
-- To change the f0 inference method to CREPE, replace `svc pre-hubert` with `svc pre-hubert -fm crepe`. You may need to reduce `--n-jobs` due to performance issues.
-- It is recommended to change the batch_size in `config.json` before the `train` command to match the VRAM capacity. The default value is optimized for Tesla T4 (16GB VRAM), but training is possible without that much VRAM.
+- It is recommended to increase the `batch_size` as much as possible in `config.json` before the `train` command to match the VRAM capacity.
+- To use `CREPE`, replace `svc pre-hubert` with `svc pre-hubert -fm crepe`.
+- To use `QuickVC`, replace `svc pre-config` with `svc pre-config -t quickvc`.
 - Silence removal and volume normalization are automatically performed (as in the upstream repo) and are not required.
 
 ### Further help
diff --git a/src/so_vits_svc_fork/__main__.py b/src/so_vits_svc_fork/__main__.py
index b2660bdb..6c22caea 100644
--- a/src/so_vits_svc_fork/__main__.py
+++ b/src/so_vits_svc_fork/__main__.py
@@ -93,7 +93,20 @@ def cli():
     type=bool,
     help="launch tensorboard",
 )
-def train(config_path: Path, model_path: Path, tensorboard: bool = False):
+@click.option(
+    "-r",
+    "--reset-optimizer",
+    default=False,
+    type=bool,
+    help="reset optimizer",
+    is_flag=True,
+)
+def train(
+    config_path: Path,
+    model_path: Path,
+    tensorboard: bool = False,
+    reset_optimizer: bool = False,
+):
     """Train model
     If D_0.pth or G_0.pth not found, automatically download from hub."""
     from .train import train
@@ -112,7 +125,9 @@ def train(config_path: Path, model_path: Path, tensorboard: bool = False):
         url = tb.launch()
         webbrowser.open(url)
 
-    train(config_path=config_path, model_path=model_path)
+    train(
+        config_path=config_path, model_path=model_path, reset_optimizer=reset_optimizer
+    )
 
 
 @cli.command()
@@ -208,7 +223,7 @@ def infer(
     device: Literal["cpu", "cuda"] = "cuda" if torch.cuda.is_available() else "cpu",
 ):
     """Inference"""
-    from .inference_main import infer
+    from so_vits_svc_fork.inference.main import infer
 
     if not auto_predict_f0:
         LOG.warning(
@@ -222,7 +237,9 @@ def infer(
     output_path = Path(output_path)
     model_path = Path(model_path)
     if model_path.is_dir():
-        model_path = list(sorted(model_path.glob("*.pth")))[-1]
+        model_path = list(
+            sorted(model_path.glob("G_*.pth"), key=lambda x: x.stat().st_mtime)
+        )[-1]
         LOG.info(f"Since model_path is a directory, use {model_path}")
     config_path = Path(config_path)
     if cluster_model_path is not None:
@@ -365,7 +382,7 @@ def vc(
     passthrough_original: bool = False,
 ) -> None:
     """Realtime inference from microphone"""
-    from .inference_main import realtime
+    from so_vits_svc_fork.inference.main import realtime
 
     if auto_predict_f0:
         LOG.warning(
@@ -381,7 +398,9 @@ def vc(
     if cluster_model_path is not None:
         cluster_model_path = Path(cluster_model_path)
     if model_path.is_dir():
-        model_path = list(sorted(model_path.glob("*.pth")))[-1]
+        model_path = list(
+            sorted(model_path.glob("G_*.pth"), key=lambda x: x.stat().st_mtime)
+        )[-1]
         LOG.info(f"Since model_path is a directory, use {model_path}")
 
     realtime(
@@ -451,7 +470,7 @@ def pre_resample(
     hop_seconds: float,
 ) -> None:
     """Preprocessing part 1: resample"""
-    from .preprocess_resample import preprocess_resample
+    from so_vits_svc_fork.preprocessing.preprocess_resample import preprocess_resample
 
     input_dir = Path(input_dir)
     output_dir = Path(output_dir)
@@ -466,6 +485,9 @@ def pre_resample(
     )
 
 
+from so_vits_svc_fork.preprocessing.preprocess_flist_config import CONFIG_TEMPLATE_DIR
+
+
 @cli.command()
 @click.option(
     "-i",
@@ -488,13 +510,21 @@ def pre_resample(
     default=Path("./configs/44k/config.json"),
     help="path to config",
 )
+@click.option(
+    "-t",
+    "--config-type",
+    type=click.Choice([x.stem for x in CONFIG_TEMPLATE_DIR.rglob("*.json")]),
+    default="so-vits-svc-4.0v1",
+    help="config type",
+)
 def pre_config(
     input_dir: Path,
     filelist_path: Path,
     config_path: Path,
+    config_type: str,
 ):
     """Preprocessing part 2: config"""
-    from .preprocess_flist_config import preprocess_config
+    from so_vits_svc_fork.preprocessing.preprocess_flist_config import preprocess_config
 
     input_dir = Path(input_dir)
     filelist_path = Path(filelist_path)
@@ -505,6 +535,7 @@ def pre_config(
         val_list_path=filelist_path / "val.txt",
         test_list_path=filelist_path / "test.txt",
         config_path=config_path,
+        config_name=config_type,
     )
 
 
@@ -527,7 +558,7 @@ def pre_config(
     "-n",
     "--n-jobs",
     type=int,
-    default=4,
+    default=None,
     help="number of jobs (optimal value may depend on your VRAM capacity and audio duration per file)",
 )
 @click.option(
@@ -552,7 +583,7 @@ def pre_hubert(
 ) -> None:
     """Preprocessing part 3: hubert
     If the HuBERT model is not found, it will be downloaded automatically."""
-    from .preprocess_hubert_f0 import preprocess_hubert_f0
+    from so_vits_svc_fork.preprocessing.preprocess_hubert_f0 import preprocess_hubert_f0
 
     input_dir = Path(input_dir)
     config_path = Path(config_path)
@@ -618,7 +649,9 @@ def pre_sd(
 
     if max_speakers == 1:
         LOG.warning("Consider using pre-split if max_speakers == 1")
-    from .preprocess_speaker_diarization import preprocess_speaker_diarization
+    from so_vits_svc_fork.preprocessing.preprocess_speaker_diarization import (
+        preprocess_speaker_diarization,
+    )
 
     preprocess_speaker_diarization(
         input_dir=input_dir,
@@ -669,7 +702,7 @@ def pre_split(
     sr: int,
 ):
     """Split audio files into multiple files"""
-    from .preprocess_split import preprocess_split
+    from so_vits_svc_fork.preprocessing.preprocess_split import preprocess_split
 
     preprocess_split(
         input_dir=input_dir,
@@ -727,6 +760,7 @@ def clean():
     help="device to use",
 )
 def onnx(input_path: Path, output_path: Path, config_path: Path, device: str) -> None:
+    raise NotImplementedError("ONNX export is not yet supported")
     """Export model to onnx"""
     input_path = Path(input_path)
     if input_path.is_dir():
@@ -738,7 +772,7 @@ def onnx(input_path: Path, output_path: Path, config_path: Path, device: str) ->
         output_path = output_path / (input_path.stem + ".onnx")
     config_path = Path(config_path)
     device_ = torch.device(device)
-    from .onnx_export import onnx_export
+    from so_vits_svc_fork.modules.onnx._export import onnx_export
 
     onnx_export(
         input_path=input_path,
diff --git a/src/so_vits_svc_fork/data_utils.py b/src/so_vits_svc_fork/data_utils.py
index 6a8caf1a..e4f15d74 100644
--- a/src/so_vits_svc_fork/data_utils.py
+++ b/src/so_vits_svc_fork/data_utils.py
@@ -6,6 +6,8 @@
 import torch.utils.data
 import torchaudio
 
+import so_vits_svc_fork.f0
+
 from . import utils
 from .modules.mel_processing import spectrogram_torch
 from .utils import load_filepaths_and_text
@@ -65,7 +67,7 @@ def get_audio(self, filename):
         spk = torch.LongTensor([self.spk_map[spk]])
 
         f0 = np.load(filename + ".f0.npy")
-        f0, uv = utils.interpolate_f0(f0)
+        f0, uv = so_vits_svc_fork.f0.interpolate_f0(f0)
         f0 = torch.FloatTensor(f0)
         uv = torch.FloatTensor(uv)
 
diff --git a/src/so_vits_svc_fork/f0.py b/src/so_vits_svc_fork/f0.py
new file mode 100644
index 00000000..255ad89e
--- /dev/null
+++ b/src/so_vits_svc_fork/f0.py
@@ -0,0 +1,237 @@
+from __future__ import annotations
+
+from logging import getLogger
+from typing import Any, Literal
+
+import numpy as np
+import torch
+import torchcrepe
+from cm_time import timer
+from numpy import dtype, float32, ndarray
+from torch import FloatTensor, Tensor
+
+LOG = getLogger(__name__)
+
+
+def normalize_f0(
+    f0: FloatTensor, x_mask: FloatTensor, uv: FloatTensor, random_scale=True
+) -> FloatTensor:
+    # calculate means based on x_mask
+    uv_sum = torch.sum(uv, dim=1, keepdim=True)
+    uv_sum[uv_sum == 0] = 9999
+    means = torch.sum(f0[:, 0, :] * uv, dim=1, keepdim=True) / uv_sum
+
+    if random_scale:
+        factor = torch.Tensor(f0.shape[0], 1).uniform_(0.8, 1.2).to(f0.device)
+    else:
+        factor = torch.ones(f0.shape[0], 1).to(f0.device)
+    # normalize f0 based on means and factor
+    f0_norm = (f0 - means.unsqueeze(-1)) * factor.unsqueeze(-1)
+    if torch.isnan(f0_norm).any():
+        exit(0)
+    return f0_norm * x_mask
+
+
+def interpolate_f0(
+    f0: ndarray[Any, dtype[float32]]
+) -> tuple[ndarray[Any, dtype[float32]], ndarray[Any, dtype[float32]]]:
+    data = np.reshape(f0, (f0.size, 1))
+
+    vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
+    vuv_vector[data > 0.0] = 1.0
+    vuv_vector[data <= 0.0] = 0.0
+
+    ip_data = data
+
+    frame_number = data.size
+    last_value = 0.0
+    for i in range(frame_number):
+        if data[i] <= 0.0:
+            j = i + 1
+            for j in range(i + 1, frame_number):
+                if data[j] > 0.0:
+                    break
+            if j < frame_number - 1:
+                if last_value > 0.0:
+                    step = (data[j] - data[i - 1]) / float(j - i)
+                    for k in range(i, j):
+                        ip_data[k] = data[i - 1] + step * (k - i + 1)
+                else:
+                    for k in range(i, j):
+                        ip_data[k] = data[j]
+            else:
+                for k in range(i, frame_number):
+                    ip_data[k] = last_value
+        else:
+            ip_data[i] = data[i]
+            last_value = data[i]
+
+    return ip_data[:, 0], vuv_vector[:, 0]
+
+
+def compute_f0_parselmouth(
+    wav_numpy: ndarray[Any, dtype[float32]],
+    p_len: None | int = None,
+    sampling_rate: int = 44100,
+    hop_length: int = 512,
+):
+    import parselmouth
+
+    x = wav_numpy
+    if p_len is None:
+        p_len = x.shape[0] // hop_length
+    else:
+        assert abs(p_len - x.shape[0] // hop_length) < 4, "pad length error"
+    time_step = hop_length / sampling_rate * 1000
+    f0_min = 50
+    f0_max = 1100
+    f0 = (
+        parselmouth.Sound(x, sampling_rate)
+        .to_pitch_ac(
+            time_step=time_step / 1000,
+            voicing_threshold=0.6,
+            pitch_floor=f0_min,
+            pitch_ceiling=f0_max,
+        )
+        .selected_array["frequency"]
+    )
+
+    pad_size = (p_len - len(f0) + 1) // 2
+    if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+        f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+    return f0
+
+
+def _resize_f0(
+    x: ndarray[Any, dtype[float32]], target_len: int
+) -> ndarray[Any, dtype[float32]]:
+    source = np.array(x)
+    source[source < 0.001] = np.nan
+    target = np.interp(
+        np.arange(0, len(source) * target_len, len(source)) / target_len,
+        np.arange(0, len(source)),
+        source,
+    )
+    res = np.nan_to_num(target)
+    return res
+
+
+def compute_f0_pyworld(
+    wav_numpy: ndarray[Any, dtype[float32]],
+    p_len: None | int = None,
+    sampling_rate: int = 44100,
+    hop_length: int = 512,
+    type_: Literal["dio", "harvest"] = "dio",
+):
+    import pyworld
+
+    if p_len is None:
+        p_len = wav_numpy.shape[0] // hop_length
+    if type_ == "dio":
+        f0, t = pyworld.dio(
+            wav_numpy.astype(np.double),
+            fs=sampling_rate,
+            f0_ceil=f0_max,
+            f0_floor=f0_min,
+            frame_period=1000 * hop_length / sampling_rate,
+        )
+    elif type_ == "harvest":
+        f0, t = pyworld.harvest(
+            wav_numpy.astype(np.double),
+            fs=sampling_rate,
+            f0_ceil=f0_max,
+            f0_floor=f0_min,
+            frame_period=1000 * hop_length / sampling_rate,
+        )
+    f0 = pyworld.stonemask(wav_numpy.astype(np.double), f0, t, sampling_rate)
+    for index, pitch in enumerate(f0):
+        f0[index] = round(pitch, 1)
+    return _resize_f0(f0, p_len)
+
+
+def compute_f0_crepe(
+    wav_numpy: ndarray[Any, dtype[float32]],
+    p_len: None | int = None,
+    sampling_rate: int = 44100,
+    hop_length: int = 512,
+    device: str = "cuda" if torch.cuda.is_available() else "cpu",
+    model: Literal["full", "tiny"] = "full",
+):
+    audio = torch.from_numpy(wav_numpy).to(device, copy=True)
+    audio = torch.unsqueeze(audio, dim=0)
+
+    if audio.ndim == 2 and audio.shape[0] > 1:
+        audio = torch.mean(audio, dim=0, keepdim=True).detach()
+    # (T) -> (1, T)
+    audio = audio.detach()
+
+    pitch: Tensor = torchcrepe.predict(
+        audio,
+        sampling_rate,
+        hop_length,
+        f0_min,
+        f0_max,
+        model,
+        batch_size=hop_length * 2,
+        device=device,
+        pad=True,
+    )
+
+    f0 = pitch.squeeze(0).cpu().numpy()
+    p_len = p_len or wav_numpy.shape[0] // hop_length
+    f0 = _resize_f0(f0, p_len)
+    return f0
+
+
+def compute_f0(
+    wav_numpy: ndarray[Any, dtype[float32]],
+    p_len: None | int = None,
+    sampling_rate: int = 44100,
+    hop_length: int = 512,
+    method: Literal["crepe", "crepe-tiny", "parselmouth", "dio", "harvest"] = "dio",
+    **kwargs,
+):
+    with timer() as t:
+        wav_numpy = wav_numpy.astype(np.float32)
+        wav_numpy /= np.quantile(np.abs(wav_numpy), 0.999)
+        if method in ["dio", "harvest"]:
+            f0 = compute_f0_pyworld(wav_numpy, p_len, sampling_rate, hop_length, method)
+        elif method == "crepe":
+            f0 = compute_f0_crepe(wav_numpy, p_len, sampling_rate, hop_length, **kwargs)
+        elif method == "crepe-tiny":
+            f0 = compute_f0_crepe(
+                wav_numpy, p_len, sampling_rate, hop_length, model="tiny", **kwargs
+            )
+        elif method == "parselmouth":
+            f0 = compute_f0_parselmouth(wav_numpy, p_len, sampling_rate, hop_length)
+        else:
+            raise ValueError(
+                "type must be dio, crepe, crepe-tiny, harvest or parselmouth"
+            )
+    rtf = t.elapsed / (len(wav_numpy) / sampling_rate)
+    LOG.info(f"F0 inference time:       {t.elapsed:.3f}s, RTF: {rtf:.3f}")
+    return f0
+
+
+def f0_to_coarse(f0: torch.Tensor | float):
+    is_torch = isinstance(f0, torch.Tensor)
+    f0_mel = 1127 * (1 + f0 / 700).log() if is_torch else 1127 * np.log(1 + f0 / 700)
+    f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * (f0_bin - 2) / (
+        f0_mel_max - f0_mel_min
+    ) + 1
+
+    f0_mel[f0_mel <= 1] = 1
+    f0_mel[f0_mel > f0_bin - 1] = f0_bin - 1
+    f0_coarse = (f0_mel + 0.5).long() if is_torch else np.rint(f0_mel).astype(np.int)
+    assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+        f0_coarse.max(),
+        f0_coarse.min(),
+    )
+    return f0_coarse
+
+
+f0_bin = 256
+f0_max = 1100.0
+f0_min = 50.0
+f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+f0_mel_max = 1127 * np.log(1 + f0_max / 700)
diff --git a/src/so_vits_svc_fork/gui.py b/src/so_vits_svc_fork/gui.py
index c70d180b..406e8ff7 100644
--- a/src/so_vits_svc_fork/gui.py
+++ b/src/so_vits_svc_fork/gui.py
@@ -11,7 +11,7 @@
 from pebble import ProcessFuture, ProcessPool
 from tqdm.tk import tqdm_tk
 
-from .utils import ensure_hubert_model
+from .utils import ensure_pretrained_model
 
 GUI_DEFAULT_PRESETS_PATH = Path(__file__).parent / "default_gui_presets.json"
 GUI_PRESETS_PATH = Path("./user_gui_presets.json").absolute()
@@ -85,16 +85,16 @@ def get_devices(
 
 def main():
     try:
-        ensure_hubert_model(tqdm_cls=tqdm_tk)
+        ensure_pretrained_model(".", "contentvec", tqdm_cls=tqdm_tk)
     except Exception as e:
         LOG.exception(e)
         LOG.info("Trying tqdm.std...")
         try:
-            ensure_hubert_model()
+            ensure_pretrained_model(".", "contentvec")
         except Exception as e:
             LOG.exception(e)
             try:
-                ensure_hubert_model(disable=True)
+                ensure_pretrained_model(".", "contentvec", disabled=True)
             except Exception as e:
                 LOG.exception(e)
                 LOG.error(
@@ -433,7 +433,7 @@ def main():
                 sg.Button("(Re)Start Voice Changer", key="start_vc"),
                 sg.Button("Stop Voice Changer", key="stop_vc"),
                 sg.Push(),
-                sg.Button("ONNX Export", key="onnx_export"),
+                # sg.Button("ONNX Export", key="onnx_export"),
             ],
         ]
     )
@@ -453,7 +453,7 @@ def update_speaker() -> None:
 
         config_path = Path(values["config_path"])
         if config_path.exists() and config_path.is_file():
-            hp = utils.get_hparams_from_file(values["config_path"])
+            hp = utils.get_hparams(values["config_path"])
             LOG.debug(f"Loaded config from {values['config_path']}")
             window["speaker"].update(
                 values=list(hp.__dict__["spk"].keys()), set_to_index=0
@@ -548,7 +548,7 @@ def apply_preset(name: str) -> None:
             elif event == "config_path":
                 update_speaker()
             elif event == "infer":
-                from .inference_main import infer
+                from so_vits_svc_fork.inference.main import infer
 
                 input_path = Path(values["input_path"])
                 output_path = (
@@ -600,7 +600,7 @@ def apply_preset(name: str) -> None:
                 _, _, input_device_indices, output_device_indices = get_devices(
                     update=False
                 )
-                from .inference_main import realtime
+                from so_vits_svc_fork.inference.main import realtime
 
                 if future:
                     LOG.info("Canceling previous task")
@@ -650,9 +650,10 @@ def apply_preset(name: str) -> None:
                     future.cancel()
                     future = None
             elif event == "onnx_export":
-                from .onnx_export import onnx_export
-
                 try:
+                    raise NotImplementedError("ONNX export is not implemented yet.")
+                    from so_vits_svc_fork.modules.onnx._export import onnx_export
+
                     onnx_export(
                         input_path=Path(values["model_path"]),
                         output_path=Path(values["model_path"]).with_suffix(".onnx"),
diff --git a/src/so_vits_svc_fork/hparams.py b/src/so_vits_svc_fork/hparams.py
new file mode 100644
index 00000000..6307042c
--- /dev/null
+++ b/src/so_vits_svc_fork/hparams.py
@@ -0,0 +1,38 @@
+from __future__ import annotations
+
+from typing import Any
+
+
+class HParams:
+    def __init__(self, **kwargs: Any) -> None:
+        for k, v in kwargs.items():
+            if type(v) == dict:
+                v = HParams(**v)
+            self[k] = v
+
+    def keys(self):
+        return self.__dict__.keys()
+
+    def items(self):
+        return self.__dict__.items()
+
+    def values(self):
+        return self.__dict__.values()
+
+    def get(self, key: str, default: Any = None):
+        return self.__dict__.get(key, default)
+
+    def __len__(self):
+        return len(self.__dict__)
+
+    def __getitem__(self, key):
+        return getattr(self, key)
+
+    def __setitem__(self, key, value):
+        return setattr(self, key, value)
+
+    def __contains__(self, key):
+        return key in self.__dict__
+
+    def __repr__(self):
+        return self.__dict__.__repr__()
diff --git a/src/so_vits_svc_fork/hubert/hubert_model.py b/src/so_vits_svc_fork/hubert/hubert_model.py
deleted file mode 100644
index 98c40d56..00000000
--- a/src/so_vits_svc_fork/hubert/hubert_model.py
+++ /dev/null
@@ -1,223 +0,0 @@
-import copy
-import random
-from typing import Optional, Tuple
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as t_func
-from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
-
-
-class Hubert(nn.Module):
-    def __init__(self, num_label_embeddings: int = 100, mask: bool = True):
-        super().__init__()
-        self._mask = mask
-        self.feature_extractor = FeatureExtractor()
-        self.feature_projection = FeatureProjection()
-        self.positional_embedding = PositionalConvEmbedding()
-        self.norm = nn.LayerNorm(768)
-        self.dropout = nn.Dropout(0.1)
-        self.encoder = TransformerEncoder(
-            nn.TransformerEncoderLayer(
-                768, 12, 3072, activation="gelu", batch_first=True
-            ),
-            12,
-        )
-        self.proj = nn.Linear(768, 256)
-
-        self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_())
-        self.label_embedding = nn.Embedding(num_label_embeddings, 256)
-
-    def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-        mask = None
-        if self.training and self._mask:
-            mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2)
-            x[mask] = self.masked_spec_embed.to(x.dtype)
-        return x, mask
-
-    def encode(
-        self, x: torch.Tensor, layer: Optional[int] = None
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-        x = self.feature_extractor(x)
-        x = self.feature_projection(x.transpose(1, 2))
-        x, mask = self.mask(x)
-        x = x + self.positional_embedding(x)
-        x = self.dropout(self.norm(x))
-        x = self.encoder(x, output_layer=layer)
-        return x, mask
-
-    def logits(self, x: torch.Tensor) -> torch.Tensor:
-        logits = torch.cosine_similarity(
-            x.unsqueeze(2),
-            self.label_embedding.weight.unsqueeze(0).unsqueeze(0),
-            dim=-1,
-        )
-        return logits / 0.1
-
-    def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-        x, mask = self.encode(x)
-        x = self.proj(x)
-        logits = self.logits(x)
-        return logits, mask
-
-
-class HubertSoft(Hubert):
-    def __init__(self):
-        super().__init__()
-
-    @torch.inference_mode()
-    def units(self, wav: torch.Tensor) -> torch.Tensor:
-        wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2))
-        x, _ = self.encode(wav)
-        return self.proj(x)
-
-
-class FeatureExtractor(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False)
-        self.norm0 = nn.GroupNorm(512, 512)
-        self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False)
-        self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False)
-        self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False)
-        self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False)
-        self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False)
-        self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        x = t_func.gelu(self.norm0(self.conv0(x)))
-        x = t_func.gelu(self.conv1(x))
-        x = t_func.gelu(self.conv2(x))
-        x = t_func.gelu(self.conv3(x))
-        x = t_func.gelu(self.conv4(x))
-        x = t_func.gelu(self.conv5(x))
-        x = t_func.gelu(self.conv6(x))
-        return x
-
-
-class FeatureProjection(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.norm = nn.LayerNorm(512)
-        self.projection = nn.Linear(512, 768)
-        self.dropout = nn.Dropout(0.1)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        x = self.norm(x)
-        x = self.projection(x)
-        x = self.dropout(x)
-        return x
-
-
-class PositionalConvEmbedding(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.conv = nn.Conv1d(
-            768,
-            768,
-            kernel_size=128,
-            padding=128 // 2,
-            groups=16,
-        )
-        self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        x = self.conv(x.transpose(1, 2))
-        x = t_func.gelu(x[:, :, :-1])
-        return x.transpose(1, 2)
-
-
-class TransformerEncoder(nn.Module):
-    def __init__(
-        self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int
-    ) -> None:
-        super().__init__()
-        self.layers = nn.ModuleList(
-            [copy.deepcopy(encoder_layer) for _ in range(num_layers)]
-        )
-        self.num_layers = num_layers
-
-    def forward(
-        self,
-        src: torch.Tensor,
-        mask: torch.Tensor = None,
-        src_key_padding_mask: torch.Tensor = None,
-        output_layer: Optional[int] = None,
-    ) -> torch.Tensor:
-        output = src
-        for layer in self.layers[:output_layer]:
-            output = layer(
-                output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
-            )
-        return output
-
-
-def _compute_mask(
-    shape: Tuple[int, int],
-    mask_prob: float,
-    mask_length: int,
-    device: torch.device,
-    min_masks: int = 0,
-) -> torch.Tensor:
-    batch_size, sequence_length = shape
-
-    if mask_length < 1:
-        raise ValueError("`mask_length` has to be bigger than 0.")
-
-    if mask_length > sequence_length:
-        raise ValueError(
-            "`mask_length` has to be smaller than `sequence_length`, "
-            f"but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`"
-        )
-
-    # compute number of masked spans in batch
-    num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random())
-    num_masked_spans = max(num_masked_spans, min_masks)
-
-    # make sure num masked indices <= sequence_length
-    if num_masked_spans * mask_length > sequence_length:
-        num_masked_spans = sequence_length // mask_length
-
-    # SpecAugment mask to fill
-    mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool)
-
-    # uniform distribution to sample from, make sure that offset samples are < sequence_length
-    uniform_dist = torch.ones(
-        (batch_size, sequence_length - (mask_length - 1)), device=device
-    )
-
-    # get random indices to mask
-    mask_indices = torch.multinomial(uniform_dist, num_masked_spans)
-
-    # expand masked indices to masked spans
-    mask_indices = (
-        mask_indices.unsqueeze(dim=-1)
-        .expand((batch_size, num_masked_spans, mask_length))
-        .reshape(batch_size, num_masked_spans * mask_length)
-    )
-    offsets = (
-        torch.arange(mask_length, device=device)[None, None, :]
-        .expand((batch_size, num_masked_spans, mask_length))
-        .reshape(batch_size, num_masked_spans * mask_length)
-    )
-    mask_idxs = mask_indices + offsets
-
-    # scatter indices to mask
-    mask = mask.scatter(1, mask_idxs, True)
-
-    return mask
-
-
-def hubert_soft(
-    path: str,
-) -> HubertSoft:
-    r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
-    Args:
-        path (str): path of a pretrained model
-    """
-    hubert = HubertSoft()
-    checkpoint = torch.load(path)
-    consume_prefix_in_state_dict_if_present(checkpoint, "module.")
-    hubert.load_state_dict(checkpoint)
-    hubert.eval()
-    return hubert
diff --git a/src/so_vits_svc_fork/hubert/hubert_model_onnx.py b/src/so_vits_svc_fork/hubert/hubert_model_onnx.py
deleted file mode 100644
index 548f599e..00000000
--- a/src/so_vits_svc_fork/hubert/hubert_model_onnx.py
+++ /dev/null
@@ -1,219 +0,0 @@
-import copy
-import random
-from typing import Optional, Tuple
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as t_func
-from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
-
-
-class Hubert(nn.Module):
-    def __init__(self, num_label_embeddings: int = 100, mask: bool = True):
-        super().__init__()
-        self._mask = mask
-        self.feature_extractor = FeatureExtractor()
-        self.feature_projection = FeatureProjection()
-        self.positional_embedding = PositionalConvEmbedding()
-        self.norm = nn.LayerNorm(768)
-        self.dropout = nn.Dropout(0.1)
-        self.encoder = TransformerEncoder(
-            nn.TransformerEncoderLayer(
-                768, 12, 3072, activation="gelu", batch_first=True
-            ),
-            12,
-        )
-        self.proj = nn.Linear(768, 256)
-
-        self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_())
-        self.label_embedding = nn.Embedding(num_label_embeddings, 256)
-
-    def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-        mask = None
-        if self.training and self._mask:
-            mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2)
-            x[mask] = self.masked_spec_embed.to(x.dtype)
-        return x, mask
-
-    def encode(
-        self, x: torch.Tensor, layer: Optional[int] = None
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-        x = self.feature_extractor(x)
-        x = self.feature_projection(x.transpose(1, 2))
-        x, mask = self.mask(x)
-        x = x + self.positional_embedding(x)
-        x = self.dropout(self.norm(x))
-        x = self.encoder(x, output_layer=layer)
-        return x, mask
-
-    def logits(self, x: torch.Tensor) -> torch.Tensor:
-        logits = torch.cosine_similarity(
-            x.unsqueeze(2),
-            self.label_embedding.weight.unsqueeze(0).unsqueeze(0),
-            dim=-1,
-        )
-        return logits / 0.1
-
-
-class HubertSoft(Hubert):
-    def __init__(self):
-        super().__init__()
-
-    def units(self, wav: torch.Tensor) -> torch.Tensor:
-        wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2))
-        x, _ = self.encode(wav)
-        return self.proj(x)
-
-    def forward(self, x):
-        return self.units(x)
-
-
-class FeatureExtractor(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False)
-        self.norm0 = nn.GroupNorm(512, 512)
-        self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False)
-        self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False)
-        self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False)
-        self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False)
-        self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False)
-        self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        x = t_func.gelu(self.norm0(self.conv0(x)))
-        x = t_func.gelu(self.conv1(x))
-        x = t_func.gelu(self.conv2(x))
-        x = t_func.gelu(self.conv3(x))
-        x = t_func.gelu(self.conv4(x))
-        x = t_func.gelu(self.conv5(x))
-        x = t_func.gelu(self.conv6(x))
-        return x
-
-
-class FeatureProjection(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.norm = nn.LayerNorm(512)
-        self.projection = nn.Linear(512, 768)
-        self.dropout = nn.Dropout(0.1)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        x = self.norm(x)
-        x = self.projection(x)
-        x = self.dropout(x)
-        return x
-
-
-class PositionalConvEmbedding(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.conv = nn.Conv1d(
-            768,
-            768,
-            kernel_size=128,
-            padding=128 // 2,
-            groups=16,
-        )
-        self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        x = self.conv(x.transpose(1, 2))
-        x = t_func.gelu(x[:, :, :-1])
-        return x.transpose(1, 2)
-
-
-class TransformerEncoder(nn.Module):
-    def __init__(
-        self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int
-    ) -> None:
-        super().__init__()
-        self.layers = nn.ModuleList(
-            [copy.deepcopy(encoder_layer) for _ in range(num_layers)]
-        )
-        self.num_layers = num_layers
-
-    def forward(
-        self,
-        src: torch.Tensor,
-        mask: torch.Tensor = None,
-        src_key_padding_mask: torch.Tensor = None,
-        output_layer: Optional[int] = None,
-    ) -> torch.Tensor:
-        output = src
-        for layer in self.layers[:output_layer]:
-            output = layer(
-                output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
-            )
-        return output
-
-
-def _compute_mask(
-    shape: Tuple[int, int],
-    mask_prob: float,
-    mask_length: int,
-    device: torch.device,
-    min_masks: int = 0,
-) -> torch.Tensor:
-    batch_size, sequence_length = shape
-
-    if mask_length < 1:
-        raise ValueError("`mask_length` has to be bigger than 0.")
-
-    if mask_length > sequence_length:
-        raise ValueError(
-            "`mask_length` has to be smaller than `sequence_length`, "
-            f"but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`"
-        )
-
-    # compute number of masked spans in batch
-    num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random())
-    num_masked_spans = max(num_masked_spans, min_masks)
-
-    # make sure num masked indices <= sequence_length
-    if num_masked_spans * mask_length > sequence_length:
-        num_masked_spans = sequence_length // mask_length
-
-    # SpecAugment mask to fill
-    mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool)
-
-    # uniform distribution to sample from, make sure that offset samples are < sequence_length
-    uniform_dist = torch.ones(
-        (batch_size, sequence_length - (mask_length - 1)), device=device
-    )
-
-    # get random indices to mask
-    mask_indices = torch.multinomial(uniform_dist, num_masked_spans)
-
-    # expand masked indices to masked spans
-    mask_indices = (
-        mask_indices.unsqueeze(dim=-1)
-        .expand((batch_size, num_masked_spans, mask_length))
-        .reshape(batch_size, num_masked_spans * mask_length)
-    )
-    offsets = (
-        torch.arange(mask_length, device=device)[None, None, :]
-        .expand((batch_size, num_masked_spans, mask_length))
-        .reshape(batch_size, num_masked_spans * mask_length)
-    )
-    mask_idxs = mask_indices + offsets
-
-    # scatter indices to mask
-    mask = mask.scatter(1, mask_idxs, True)
-
-    return mask
-
-
-def hubert_soft(
-    path: str,
-) -> HubertSoft:
-    r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
-    Args:
-        path (str): path of a pretrained model
-    """
-    hubert = HubertSoft()
-    checkpoint = torch.load(path)
-    consume_prefix_in_state_dict_if_present(checkpoint, "module.")
-    hubert.load_state_dict(checkpoint)
-    hubert.eval()
-    return hubert
diff --git a/src/so_vits_svc_fork/hubert/put_hubert_ckpt_here b/src/so_vits_svc_fork/hubert/put_hubert_ckpt_here
deleted file mode 100644
index e69de29b..00000000
diff --git a/src/so_vits_svc_fork/inference/infer_tool.py b/src/so_vits_svc_fork/inference/core.py
similarity index 91%
rename from src/so_vits_svc_fork/inference/infer_tool.py
rename to src/so_vits_svc_fork/inference/core.py
index 5604192f..fbbcd794 100644
--- a/src/so_vits_svc_fork/inference/infer_tool.py
+++ b/src/so_vits_svc_fork/inference/core.py
@@ -12,10 +12,10 @@
 from cm_time import timer
 from numpy import dtype, float32, ndarray
 
+import so_vits_svc_fork.f0
 from so_vits_svc_fork import cluster, utils
-from so_vits_svc_fork.models import SynthesizerTrn
 
-from ..utils import HUBERT_SAMPLING_RATE
+from ..modules.synthesizers import SynthesizerTrn
 
 LOG = getLogger(__name__)
 
@@ -90,37 +90,38 @@ class Svc:
     def __init__(
         self,
         *,
-        net_g_path: str,
-        config_path: str,
+        net_g_path: Path | str,
+        config_path: Path | str,
         device: torch.device | str | None = None,
         cluster_model_path: Path | str | None = None,
+        half: bool = False,
     ):
         self.net_g_path = net_g_path
         if device is None:
-            self.dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         else:
-            self.dev = torch.device(device)
-        self.net_g_ms = None
-        self.hps_ms = utils.get_hparams_from_file(config_path)
-        self.target_sample = self.hps_ms.data.sampling_rate
-        self.hop_size = self.hps_ms.data.hop_length
-        self.spk2id = self.hps_ms.spk
-        self.hubert_model = utils.get_hubert_model().to(self.dev)
+            self.device = torch.device(device)
+        self.hps = utils.get_hparams(config_path)
+        self.target_sample = self.hps.data.sampling_rate
+        self.hop_size = self.hps.data.hop_length
+        self.spk2id = self.hps.spk
+        self.hubert_model = utils.get_hubert_model(self.device)
+        self.dtype = torch.float16 if half else torch.float32
+        self.contentvec_final_proj = self.hps.data.__dict__.get(
+            "contentvec_final_proj", True
+        )
         self.load_model()
         if cluster_model_path is not None and Path(cluster_model_path).exists():
             self.cluster_model = cluster.get_cluster_model(cluster_model_path)
 
     def load_model(self):
-        self.net_g_ms = SynthesizerTrn(
-            self.hps_ms.data.filter_length // 2 + 1,
-            self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
-            **self.hps_ms.model,
+        self.net_g = SynthesizerTrn(
+            self.hps.data.filter_length // 2 + 1,
+            self.hps.train.segment_size // self.hps.data.hop_length,
+            **self.hps.model,
         )
-        _ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None)
-        if "half" in self.net_g_path and torch.cuda.is_available():
-            _ = self.net_g_ms.half().eval().to(self.dev)
-        else:
-            _ = self.net_g_ms.eval().to(self.dev)
+        _ = utils.load_checkpoint(self.net_g_path, self.net_g, None)
+        _ = self.net_g.eval().to(self.device, dtype=self.dtype)
 
     def get_unit_f0(
         self,
@@ -132,31 +133,33 @@ def get_unit_f0(
             "crepe", "crepe-tiny", "parselmouth", "dio", "harvest"
         ] = "dio",
     ):
-        f0 = utils.compute_f0(
+        f0 = so_vits_svc_fork.f0.compute_f0(
             audio,
             sampling_rate=self.target_sample,
             hop_length=self.hop_size,
             method=f0_method,
         )
-        f0, uv = utils.interpolate_f0(f0)
-        f0 = torch.FloatTensor(f0)
-        uv = torch.FloatTensor(uv)
+        f0, uv = so_vits_svc_fork.f0.interpolate_f0(f0)
+        f0 = torch.as_tensor(f0, dtype=self.dtype, device=self.device)
+        uv = torch.as_tensor(uv, dtype=self.dtype, device=self.device)
         f0 = f0 * 2 ** (tran / 12)
-        f0 = f0.unsqueeze(0).to(self.dev)
-        uv = uv.unsqueeze(0).to(self.dev)
+        f0 = f0.unsqueeze(0)
+        uv = uv.unsqueeze(0)
 
-        wav16k = librosa.resample(
-            audio, orig_sr=self.target_sample, target_sr=HUBERT_SAMPLING_RATE
-        )
-        wav16k = torch.from_numpy(wav16k).to(self.dev)
-        c = utils.get_hubert_content(self.hubert_model, wav_16k_tensor=wav16k)
+        c = utils.get_content(
+            self.hubert_model,
+            audio,
+            self.device,
+            self.target_sample,
+            self.contentvec_final_proj,
+        ).to(self.dtype)
         c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1])
 
         if cluster_infer_ratio != 0:
             cluster_c = cluster.get_cluster_center_result(
                 self.cluster_model, c.cpu().numpy().T, speaker
             ).T
-            cluster_c = torch.FloatTensor(cluster_c).to(self.dev)
+            cluster_c = torch.FloatTensor(cluster_c).to(self.device)
             c = cluster_infer_ratio * cluster_c + (1 - cluster_infer_ratio) * c
 
         c = c.unsqueeze(0)
@@ -199,19 +202,17 @@ def infer(
         elif len(speaker_candidates) == 0:
             raise ValueError(f"Speaker_id {speaker_id} is not found.")
         speaker = speaker_candidates[0][0]
-        sid = torch.LongTensor([int(speaker_id)]).to(self.dev).unsqueeze(0)
+        sid = torch.LongTensor([int(speaker_id)]).to(self.device).unsqueeze(0)
 
         # get unit f0
         c, f0, uv = self.get_unit_f0(
             audio, transpose, cluster_infer_ratio, speaker, f0_method
         )
-        if "half" in self.net_g_path and torch.cuda.is_available():
-            c = c.half()
 
         # inference
         with torch.no_grad():
             with timer() as t:
-                audio = self.net_g_ms.infer(
+                audio = self.net_g.infer(
                     c,
                     f0=f0,
                     g=sid,
@@ -223,10 +224,8 @@ def infer(
             LOG.info(
                 f"Inferece time: {t.elapsed:.2f}s, RTF: {t.elapsed / audio_duration:.2f}"
             )
-        return audio, audio.shape[-1]
-
-    def clear_empty(self):
         torch.cuda.empty_cache()
+        return audio, audio.shape[-1]
 
     def infer_silence(
         self,
@@ -253,7 +252,7 @@ def infer_silence(
         chunk_length_min = chunk_length_min = (
             int(
                 min(
-                    sr / utils.f0_min * 20 + 1,
+                    sr / so_vits_svc_fork.f0.f0_min * 20 + 1,
                     chunk_seconds * sr,
                 )
             )
@@ -299,6 +298,9 @@ def infer_silence(
                 # fade_len = int(self.target_sample * fade_seconds)
                 # _audio[:fade_len] = _audio[:fade_len] * np.linspace(0, 1, fade_len)
                 # _audio[-fade_len:] = _audio[-fade_len:] * np.linspace(1, 0, fade_len)
+
+                # empty cache
+                torch.cuda.empty_cache()
             result_audio = np.concatenate([result_audio, audio_chunk_infer])
         result_audio = result_audio[: audio.shape[0]]
         return result_audio
@@ -539,7 +541,9 @@ def infer(audio: ndarray[Any, dtype[float32]]) -> ndarray[Any, dtype[float32]]:
         self.input_audio_store = np.concatenate([self.input_audio_store, input_audio])
         LOG.info(f"input_audio_store: {self.input_audio_store.shape}")
         sr = self.svc_model.target_sample
-        chunk_length_min = int(min(sr / utils.f0_min * 20 + 1, chunk_seconds * sr)) // 2
+        chunk_length_min = (
+            int(min(sr / so_vits_svc_fork.f0.f0_min * 20 + 1, chunk_seconds * sr)) // 2
+        )
         LOG.info(f"Chunk length min: {chunk_length_min}")
         chunk_list = list(
             split_silence(
diff --git a/src/so_vits_svc_fork/inference_main.py b/src/so_vits_svc_fork/inference/main.py
similarity index 99%
rename from src/so_vits_svc_fork/inference_main.py
rename to src/so_vits_svc_fork/inference/main.py
index e01ed63d..85852b26 100644
--- a/src/so_vits_svc_fork/inference_main.py
+++ b/src/so_vits_svc_fork/inference/main.py
@@ -10,7 +10,7 @@
 import torch
 from cm_time import timer
 
-from .inference.infer_tool import RealtimeVC, RealtimeVC2, Svc
+from so_vits_svc_fork.inference.core import RealtimeVC, RealtimeVC2, Svc
 
 LOG = getLogger(__name__)
 
diff --git a/src/so_vits_svc_fork/models.py b/src/so_vits_svc_fork/models.py
deleted file mode 100644
index 2dc3db7c..00000000
--- a/src/so_vits_svc_fork/models.py
+++ /dev/null
@@ -1,519 +0,0 @@
-import torch
-from torch import nn
-from torch.nn import Conv1d, Conv2d
-from torch.nn import functional as F
-from torch.nn.utils import spectral_norm, weight_norm
-
-import so_vits_svc_fork.modules.attentions as attentions
-import so_vits_svc_fork.modules.commons as commons
-import so_vits_svc_fork.modules.modules as modules
-
-from . import utils
-from .modules.commons import get_padding
-from .utils import f0_to_coarse
-from .vdecoder.hifigan.models import Generator
-
-
-class ResidualCouplingBlock(nn.Module):
-    def __init__(
-        self,
-        channels,
-        hidden_channels,
-        kernel_size,
-        dilation_rate,
-        n_layers,
-        n_flows=4,
-        gin_channels=0,
-    ):
-        super().__init__()
-        self.channels = channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.dilation_rate = dilation_rate
-        self.n_layers = n_layers
-        self.n_flows = n_flows
-        self.gin_channels = gin_channels
-
-        self.flows = nn.ModuleList()
-        for i in range(n_flows):
-            self.flows.append(
-                modules.ResidualCouplingLayer(
-                    channels,
-                    hidden_channels,
-                    kernel_size,
-                    dilation_rate,
-                    n_layers,
-                    gin_channels=gin_channels,
-                    mean_only=True,
-                )
-            )
-            self.flows.append(modules.Flip())
-
-    def forward(self, x, x_mask, g=None, reverse=False):
-        if not reverse:
-            for flow in self.flows:
-                x, _ = flow(x, x_mask, g=g, reverse=reverse)
-        else:
-            for flow in reversed(self.flows):
-                x = flow(x, x_mask, g=g, reverse=reverse)
-        return x
-
-
-class Encoder(nn.Module):
-    def __init__(
-        self,
-        in_channels,
-        out_channels,
-        hidden_channels,
-        kernel_size,
-        dilation_rate,
-        n_layers,
-        gin_channels=0,
-    ):
-        super().__init__()
-        self.in_channels = in_channels
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.dilation_rate = dilation_rate
-        self.n_layers = n_layers
-        self.gin_channels = gin_channels
-
-        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
-        self.enc = modules.WN(
-            hidden_channels,
-            kernel_size,
-            dilation_rate,
-            n_layers,
-            gin_channels=gin_channels,
-        )
-        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
-
-    def forward(self, x, x_lengths, g=None):
-        # print(x.shape,x_lengths.shape)
-        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
-            x.dtype
-        )
-        x = self.pre(x) * x_mask
-        x = self.enc(x, x_mask, g=g)
-        stats = self.proj(x) * x_mask
-        m, logs = torch.split(stats, self.out_channels, dim=1)
-        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
-        return z, m, logs, x_mask
-
-
-class TextEncoder(nn.Module):
-    def __init__(
-        self,
-        out_channels,
-        hidden_channels,
-        kernel_size,
-        n_layers,
-        gin_channels=0,
-        filter_channels=None,
-        n_heads=None,
-        p_dropout=None,
-    ):
-        super().__init__()
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.n_layers = n_layers
-        self.gin_channels = gin_channels
-        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
-        self.f0_emb = nn.Embedding(256, hidden_channels)
-
-        self.enc_ = attentions.Encoder(
-            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
-        )
-
-    def forward(self, x, x_mask, f0=None, noice_scale=1):
-        x = x + self.f0_emb(f0).transpose(1, 2)
-        x = self.enc_(x * x_mask, x_mask)
-        stats = self.proj(x) * x_mask
-        m, logs = torch.split(stats, self.out_channels, dim=1)
-        z = (m + torch.randn_like(m) * torch.exp(logs) * noice_scale) * x_mask
-
-        return z, m, logs, x_mask
-
-
-class DiscriminatorP(torch.nn.Module):
-    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
-        super().__init__()
-        self.period = period
-        self.use_spectral_norm = use_spectral_norm
-        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
-        self.convs = nn.ModuleList(
-            [
-                norm_f(
-                    Conv2d(
-                        1,
-                        32,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        32,
-                        128,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        128,
-                        512,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        512,
-                        1024,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        1024,
-                        1024,
-                        (kernel_size, 1),
-                        1,
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-            ]
-        )
-        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
-
-    def forward(self, x):
-        fmap = []
-
-        # 1d to 2d
-        b, c, t = x.shape
-        if t % self.period != 0:  # pad first
-            n_pad = self.period - (t % self.period)
-            x = F.pad(x, (0, n_pad), "reflect")
-            t = t + n_pad
-        x = x.view(b, c, t // self.period, self.period)
-
-        for l in self.convs:
-            x = l(x)
-            x = F.leaky_relu(x, modules.LRELU_SLOPE)
-            fmap.append(x)
-        x = self.conv_post(x)
-        fmap.append(x)
-        x = torch.flatten(x, 1, -1)
-
-        return x, fmap
-
-
-class DiscriminatorS(torch.nn.Module):
-    def __init__(self, use_spectral_norm=False):
-        super().__init__()
-        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
-        self.convs = nn.ModuleList(
-            [
-                norm_f(Conv1d(1, 16, 15, 1, padding=7)),
-                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
-                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
-                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
-                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
-                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
-            ]
-        )
-        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
-
-    def forward(self, x):
-        fmap = []
-
-        for l in self.convs:
-            x = l(x)
-            x = F.leaky_relu(x, modules.LRELU_SLOPE)
-            fmap.append(x)
-        x = self.conv_post(x)
-        fmap.append(x)
-        x = torch.flatten(x, 1, -1)
-
-        return x, fmap
-
-
-class MultiPeriodDiscriminator(torch.nn.Module):
-    def __init__(self, use_spectral_norm=False):
-        super().__init__()
-        periods = [2, 3, 5, 7, 11]
-
-        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
-        discs = discs + [
-            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
-        ]
-        self.discriminators = nn.ModuleList(discs)
-
-    def forward(self, y, y_hat):
-        y_d_rs = []
-        y_d_gs = []
-        fmap_rs = []
-        fmap_gs = []
-        for i, d in enumerate(self.discriminators):
-            y_d_r, fmap_r = d(y)
-            y_d_g, fmap_g = d(y_hat)
-            y_d_rs.append(y_d_r)
-            y_d_gs.append(y_d_g)
-            fmap_rs.append(fmap_r)
-            fmap_gs.append(fmap_g)
-
-        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
-
-
-class SpeakerEncoder(torch.nn.Module):
-    def __init__(
-        self,
-        mel_n_channels=80,
-        model_num_layers=3,
-        model_hidden_size=256,
-        model_embedding_size=256,
-    ):
-        super().__init__()
-        self.lstm = nn.LSTM(
-            mel_n_channels, model_hidden_size, model_num_layers, batch_first=True
-        )
-        self.linear = nn.Linear(model_hidden_size, model_embedding_size)
-        self.relu = nn.ReLU()
-
-    def forward(self, mels):
-        self.lstm.flatten_parameters()
-        _, (hidden, _) = self.lstm(mels)
-        embeds_raw = self.relu(self.linear(hidden[-1]))
-        return embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True)
-
-    def compute_partial_slices(self, total_frames, partial_frames, partial_hop):
-        mel_slices = []
-        for i in range(0, total_frames - partial_frames, partial_hop):
-            mel_range = torch.arange(i, i + partial_frames)
-            mel_slices.append(mel_range)
-
-        return mel_slices
-
-    def embed_utterance(self, mel, partial_frames=128, partial_hop=64):
-        mel_len = mel.size(1)
-        last_mel = mel[:, -partial_frames:]
-
-        if mel_len > partial_frames:
-            mel_slices = self.compute_partial_slices(
-                mel_len, partial_frames, partial_hop
-            )
-            mels = list(mel[:, s] for s in mel_slices)
-            mels.append(last_mel)
-            mels = torch.stack(tuple(mels), 0).squeeze(1)
-
-            with torch.no_grad():
-                partial_embeds = self(mels)
-            embed = torch.mean(partial_embeds, axis=0).unsqueeze(0)
-            # embed = embed / torch.linalg.norm(embed, 2)
-        else:
-            with torch.no_grad():
-                embed = self(last_mel)
-
-        return embed
-
-
-class F0Decoder(nn.Module):
-    def __init__(
-        self,
-        out_channels,
-        hidden_channels,
-        filter_channels,
-        n_heads,
-        n_layers,
-        kernel_size,
-        p_dropout,
-        spk_channels=0,
-    ):
-        super().__init__()
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.filter_channels = filter_channels
-        self.n_heads = n_heads
-        self.n_layers = n_layers
-        self.kernel_size = kernel_size
-        self.p_dropout = p_dropout
-        self.spk_channels = spk_channels
-
-        self.prenet = nn.Conv1d(hidden_channels, hidden_channels, 3, padding=1)
-        self.decoder = attentions.FFT(
-            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
-        )
-        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
-        self.f0_prenet = nn.Conv1d(1, hidden_channels, 3, padding=1)
-        self.cond = nn.Conv1d(spk_channels, hidden_channels, 1)
-
-    def forward(self, x, norm_f0, x_mask, spk_emb=None):
-        x = torch.detach(x)
-        if spk_emb is not None:
-            x = x + self.cond(spk_emb)
-        x += self.f0_prenet(norm_f0)
-        x = self.prenet(x) * x_mask
-        x = self.decoder(x * x_mask, x_mask)
-        x = self.proj(x) * x_mask
-        return x
-
-
-class SynthesizerTrn(nn.Module):
-    """
-    Synthesizer for Training
-    """
-
-    def __init__(
-        self,
-        spec_channels,
-        segment_size,
-        inter_channels,
-        hidden_channels,
-        filter_channels,
-        n_heads,
-        n_layers,
-        kernel_size,
-        p_dropout,
-        resblock,
-        resblock_kernel_sizes,
-        resblock_dilation_sizes,
-        upsample_rates,
-        upsample_initial_channel,
-        upsample_kernel_sizes,
-        gin_channels,
-        ssl_dim,
-        n_speakers,
-        sampling_rate=44100,
-        **kwargs
-    ):
-        super().__init__()
-        self.spec_channels = spec_channels
-        self.inter_channels = inter_channels
-        self.hidden_channels = hidden_channels
-        self.filter_channels = filter_channels
-        self.n_heads = n_heads
-        self.n_layers = n_layers
-        self.kernel_size = kernel_size
-        self.p_dropout = p_dropout
-        self.resblock = resblock
-        self.resblock_kernel_sizes = resblock_kernel_sizes
-        self.resblock_dilation_sizes = resblock_dilation_sizes
-        self.upsample_rates = upsample_rates
-        self.upsample_initial_channel = upsample_initial_channel
-        self.upsample_kernel_sizes = upsample_kernel_sizes
-        self.segment_size = segment_size
-        self.gin_channels = gin_channels
-        self.ssl_dim = ssl_dim
-        self.emb_g = nn.Embedding(n_speakers, gin_channels)
-
-        self.pre = nn.Conv1d(ssl_dim, hidden_channels, kernel_size=5, padding=2)
-
-        self.enc_p = TextEncoder(
-            inter_channels,
-            hidden_channels,
-            filter_channels=filter_channels,
-            n_heads=n_heads,
-            n_layers=n_layers,
-            kernel_size=kernel_size,
-            p_dropout=p_dropout,
-        )
-        hps = {
-            "sampling_rate": sampling_rate,
-            "inter_channels": inter_channels,
-            "resblock": resblock,
-            "resblock_kernel_sizes": resblock_kernel_sizes,
-            "resblock_dilation_sizes": resblock_dilation_sizes,
-            "upsample_rates": upsample_rates,
-            "upsample_initial_channel": upsample_initial_channel,
-            "upsample_kernel_sizes": upsample_kernel_sizes,
-            "gin_channels": gin_channels,
-        }
-        self.dec = Generator(h=hps)
-        self.enc_q = Encoder(
-            spec_channels,
-            inter_channels,
-            hidden_channels,
-            5,
-            1,
-            16,
-            gin_channels=gin_channels,
-        )
-        self.flow = ResidualCouplingBlock(
-            inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels
-        )
-        self.f0_decoder = F0Decoder(
-            1,
-            hidden_channels,
-            filter_channels,
-            n_heads,
-            n_layers,
-            kernel_size,
-            p_dropout,
-            spk_channels=gin_channels,
-        )
-        self.emb_uv = nn.Embedding(2, hidden_channels)
-
-    def forward(self, c, f0, uv, spec, g=None, c_lengths=None, spec_lengths=None):
-        g = self.emb_g(g).transpose(1, 2)
-        # ssl prenet
-        x_mask = torch.unsqueeze(commons.sequence_mask(c_lengths, c.size(2)), 1).to(
-            c.dtype
-        )
-        x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1, 2)
-
-        # f0 predict
-        lf0 = 2595.0 * torch.log10(1.0 + f0.unsqueeze(1) / 700.0) / 500
-        norm_lf0 = utils.normalize_f0(lf0, x_mask, uv)
-        pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
-
-        # encoder
-        z_ptemp, m_p, logs_p, _ = self.enc_p(x, x_mask, f0=f0_to_coarse(f0))
-        z, m_q, logs_q, spec_mask = self.enc_q(spec, spec_lengths, g=g)
-
-        # flow
-        z_p = self.flow(z, spec_mask, g=g)
-        z_slice, pitch_slice, ids_slice = commons.rand_slice_segments_with_pitch(
-            z, f0, spec_lengths, self.segment_size
-        )
-
-        # nsf decoder
-        o = self.dec(z_slice, g=g, f0=pitch_slice)
-
-        return (
-            o,
-            ids_slice,
-            spec_mask,
-            (z, z_p, m_p, logs_p, m_q, logs_q),
-            pred_lf0,
-            norm_lf0,
-            lf0,
-        )
-
-    def infer(self, c, f0, uv, g=None, noice_scale=0.35, predict_f0=False):
-        c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device)
-        g = self.emb_g(g).transpose(1, 2)
-        x_mask = torch.unsqueeze(commons.sequence_mask(c_lengths, c.size(2)), 1).to(
-            c.dtype
-        )
-        x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1, 2)
-
-        if predict_f0:
-            lf0 = 2595.0 * torch.log10(1.0 + f0.unsqueeze(1) / 700.0) / 500
-            norm_lf0 = utils.normalize_f0(lf0, x_mask, uv, random_scale=False)
-            pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
-            f0 = (700 * (torch.pow(10, pred_lf0 * 500 / 2595) - 1)).squeeze(1)
-
-        z_p, m_p, logs_p, c_mask = self.enc_p(
-            x, x_mask, f0=f0_to_coarse(f0), noice_scale=noice_scale
-        )
-        z = self.flow(z_p, c_mask, g=g, reverse=True)
-        o = self.dec(z * c_mask, g=g, f0=f0)
-        return o
diff --git a/src/so_vits_svc_fork/hubert/__init__.py b/src/so_vits_svc_fork/modules/decoders/__init__.py
similarity index 100%
rename from src/so_vits_svc_fork/hubert/__init__.py
rename to src/so_vits_svc_fork/modules/decoders/__init__.py
diff --git a/src/so_vits_svc_fork/modules/decoders/f0.py b/src/so_vits_svc_fork/modules/decoders/f0.py
new file mode 100644
index 00000000..38d8c77d
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/decoders/f0.py
@@ -0,0 +1,45 @@
+import torch
+from torch import nn
+
+from so_vits_svc_fork.modules import attentions as attentions
+
+
+class F0Decoder(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        spk_channels=0,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.spk_channels = spk_channels
+
+        self.prenet = nn.Conv1d(hidden_channels, hidden_channels, 3, padding=1)
+        self.decoder = attentions.FFT(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+        self.f0_prenet = nn.Conv1d(1, hidden_channels, 3, padding=1)
+        self.cond = nn.Conv1d(spk_channels, hidden_channels, 1)
+
+    def forward(self, x, norm_f0, x_mask, spk_emb=None):
+        x = torch.detach(x)
+        if spk_emb is not None:
+            x = x + self.cond(spk_emb)
+        x += self.f0_prenet(norm_f0)
+        x = self.prenet(x) * x_mask
+        x = self.decoder(x * x_mask, x_mask)
+        x = self.proj(x) * x_mask
+        return x
diff --git a/src/so_vits_svc_fork/modules/decoders/hifigan/__init__.py b/src/so_vits_svc_fork/modules/decoders/hifigan/__init__.py
new file mode 100644
index 00000000..486c66fe
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/decoders/hifigan/__init__.py
@@ -0,0 +1,3 @@
+from ._models import Generator
+
+__all__ = ["Generator"]
diff --git a/src/so_vits_svc_fork/vdecoder/hifigan/models.py b/src/so_vits_svc_fork/modules/decoders/hifigan/_models.py
similarity index 50%
rename from src/so_vits_svc_fork/vdecoder/hifigan/models.py
rename to src/so_vits_svc_fork/modules/decoders/hifigan/_models.py
index 1c83dea9..5161ed2e 100644
--- a/src/so_vits_svc_fork/vdecoder/hifigan/models.py
+++ b/src/so_vits_svc_fork/modules/decoders/hifigan/_models.py
@@ -1,175 +1,20 @@
-import json
-import os
 from logging import getLogger
 
 import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from torch.nn import AvgPool1d, Conv1d, Conv2d, ConvTranspose1d
-from torch.nn.utils import remove_weight_norm, spectral_norm, weight_norm
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils import remove_weight_norm, weight_norm
 
-from .env import AttrDict
-from .utils import get_padding, init_weights
+from ...modules import ResBlock1, ResBlock2
+from ._utils import init_weights
 
 LOG = getLogger(__name__)
 
 LRELU_SLOPE = 0.1
 
 
-def load_model(model_path, device="cuda"):
-    config_file = os.path.join(os.path.split(model_path)[0], "config.json")
-    with open(config_file) as f:
-        data = f.read()
-
-    global h
-    json_config = json.loads(data)
-    h = AttrDict(json_config)
-
-    generator = Generator(h).to(device)
-
-    cp_dict = torch.load(model_path)
-    generator.load_state_dict(cp_dict["generator"])
-    generator.eval()
-    generator.remove_weight_norm()
-    del cp_dict
-    return generator, h
-
-
-class ResBlock1(torch.nn.Module):
-    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
-        super().__init__()
-        self.h = h
-        self.convs1 = nn.ModuleList(
-            [
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[0],
-                        padding=get_padding(kernel_size, dilation[0]),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[1],
-                        padding=get_padding(kernel_size, dilation[1]),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[2],
-                        padding=get_padding(kernel_size, dilation[2]),
-                    )
-                ),
-            ]
-        )
-        self.convs1.apply(init_weights)
-
-        self.convs2 = nn.ModuleList(
-            [
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=1,
-                        padding=get_padding(kernel_size, 1),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=1,
-                        padding=get_padding(kernel_size, 1),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=1,
-                        padding=get_padding(kernel_size, 1),
-                    )
-                ),
-            ]
-        )
-        self.convs2.apply(init_weights)
-
-    def forward(self, x):
-        for c1, c2 in zip(self.convs1, self.convs2):
-            xt = F.leaky_relu(x, LRELU_SLOPE)
-            xt = c1(xt)
-            xt = F.leaky_relu(xt, LRELU_SLOPE)
-            xt = c2(xt)
-            x = xt + x
-        return x
-
-    def remove_weight_norm(self):
-        for l in self.convs1:
-            remove_weight_norm(l)
-        for l in self.convs2:
-            remove_weight_norm(l)
-
-
-class ResBlock2(torch.nn.Module):
-    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3)):
-        super().__init__()
-        self.h = h
-        self.convs = nn.ModuleList(
-            [
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[0],
-                        padding=get_padding(kernel_size, dilation[0]),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[1],
-                        padding=get_padding(kernel_size, dilation[1]),
-                    )
-                ),
-            ]
-        )
-        self.convs.apply(init_weights)
-
-    def forward(self, x):
-        for c in self.convs:
-            xt = F.leaky_relu(x, LRELU_SLOPE)
-            xt = c(xt)
-            x = xt + x
-        return x
-
-    def remove_weight_norm(self):
-        for l in self.convs:
-            remove_weight_norm(l)
-
-
 def padDiff(x):
     return F.pad(
         F.pad(x, (0, 0, -1, 1), "constant", 0) - x, (0, 0, 0, -1), "constant", 0
@@ -286,8 +131,11 @@ def forward(self, f0):
         with torch.no_grad():
             # f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
             # fundamental component
+            # fn = torch.multiply(
+            #    f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device)
+            # )
             fn = torch.multiply(
-                f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device)
+                f0, torch.arange(1, self.harmonic_num + 2).to(f0.device).to(f0.dtype)
             )
 
             # generate sine waveforms
@@ -417,7 +265,7 @@ def __init__(self, h):
             for j, (k, d) in enumerate(
                 zip(h["resblock_kernel_sizes"], h["resblock_dilation_sizes"])
             ):
-                self.resblocks.append(resblock(h, ch, k, d))
+                self.resblocks.append(resblock(ch, k, d))
 
         self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
         self.ups.apply(init_weights)
@@ -461,194 +309,3 @@ def remove_weight_norm(self):
             l.remove_weight_norm()
         remove_weight_norm(self.conv_pre)
         remove_weight_norm(self.conv_post)
-
-
-class DiscriminatorP(torch.nn.Module):
-    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
-        super().__init__()
-        self.period = period
-        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
-        self.convs = nn.ModuleList(
-            [
-                norm_f(
-                    Conv2d(
-                        1,
-                        32,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(5, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        32,
-                        128,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(5, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        128,
-                        512,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(5, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        512,
-                        1024,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(5, 1), 0),
-                    )
-                ),
-                norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
-            ]
-        )
-        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
-
-    def forward(self, x):
-        fmap = []
-
-        # 1d to 2d
-        b, c, t = x.shape
-        if t % self.period != 0:  # pad first
-            n_pad = self.period - (t % self.period)
-            x = F.pad(x, (0, n_pad), "reflect")
-            t = t + n_pad
-        x = x.view(b, c, t // self.period, self.period)
-
-        for l in self.convs:
-            x = l(x)
-            x = F.leaky_relu(x, LRELU_SLOPE)
-            fmap.append(x)
-        x = self.conv_post(x)
-        fmap.append(x)
-        x = torch.flatten(x, 1, -1)
-
-        return x, fmap
-
-
-class MultiPeriodDiscriminator(torch.nn.Module):
-    def __init__(self, periods=None):
-        super().__init__()
-        self.periods = periods if periods is not None else [2, 3, 5, 7, 11]
-        self.discriminators = nn.ModuleList()
-        for period in self.periods:
-            self.discriminators.append(DiscriminatorP(period))
-
-    def forward(self, y, y_hat):
-        y_d_rs = []
-        y_d_gs = []
-        fmap_rs = []
-        fmap_gs = []
-        for i, d in enumerate(self.discriminators):
-            y_d_r, fmap_r = d(y)
-            y_d_g, fmap_g = d(y_hat)
-            y_d_rs.append(y_d_r)
-            fmap_rs.append(fmap_r)
-            y_d_gs.append(y_d_g)
-            fmap_gs.append(fmap_g)
-
-        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
-
-
-class DiscriminatorS(torch.nn.Module):
-    def __init__(self, use_spectral_norm=False):
-        super().__init__()
-        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
-        self.convs = nn.ModuleList(
-            [
-                norm_f(Conv1d(1, 128, 15, 1, padding=7)),
-                norm_f(Conv1d(128, 128, 41, 2, groups=4, padding=20)),
-                norm_f(Conv1d(128, 256, 41, 2, groups=16, padding=20)),
-                norm_f(Conv1d(256, 512, 41, 4, groups=16, padding=20)),
-                norm_f(Conv1d(512, 1024, 41, 4, groups=16, padding=20)),
-                norm_f(Conv1d(1024, 1024, 41, 1, groups=16, padding=20)),
-                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
-            ]
-        )
-        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
-
-    def forward(self, x):
-        fmap = []
-        for l in self.convs:
-            x = l(x)
-            x = F.leaky_relu(x, LRELU_SLOPE)
-            fmap.append(x)
-        x = self.conv_post(x)
-        fmap.append(x)
-        x = torch.flatten(x, 1, -1)
-
-        return x, fmap
-
-
-class MultiScaleDiscriminator(torch.nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.discriminators = nn.ModuleList(
-            [
-                DiscriminatorS(use_spectral_norm=True),
-                DiscriminatorS(),
-                DiscriminatorS(),
-            ]
-        )
-        self.meanpools = nn.ModuleList(
-            [AvgPool1d(4, 2, padding=2), AvgPool1d(4, 2, padding=2)]
-        )
-
-    def forward(self, y, y_hat):
-        y_d_rs = []
-        y_d_gs = []
-        fmap_rs = []
-        fmap_gs = []
-        for i, d in enumerate(self.discriminators):
-            if i != 0:
-                y = self.meanpools[i - 1](y)
-                y_hat = self.meanpools[i - 1](y_hat)
-            y_d_r, fmap_r = d(y)
-            y_d_g, fmap_g = d(y_hat)
-            y_d_rs.append(y_d_r)
-            fmap_rs.append(fmap_r)
-            y_d_gs.append(y_d_g)
-            fmap_gs.append(fmap_g)
-
-        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
-
-
-def feature_loss(fmap_r, fmap_g):
-    loss = 0
-    for dr, dg in zip(fmap_r, fmap_g):
-        for rl, gl in zip(dr, dg):
-            loss += torch.mean(torch.abs(rl - gl))
-
-    return loss * 2
-
-
-def discriminator_loss(disc_real_outputs, disc_generated_outputs):
-    loss = 0
-    r_losses = []
-    g_losses = []
-    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
-        r_loss = torch.mean((1 - dr) ** 2)
-        g_loss = torch.mean(dg**2)
-        loss += r_loss + g_loss
-        r_losses.append(r_loss.item())
-        g_losses.append(g_loss.item())
-
-    return loss, r_losses, g_losses
-
-
-def generator_loss(disc_outputs):
-    loss = 0
-    gen_losses = []
-    for dg in disc_outputs:
-        l = torch.mean((1 - dg) ** 2)
-        gen_losses.append(l)
-        loss += l
-
-    return loss, gen_losses
diff --git a/src/so_vits_svc_fork/modules/decoders/hifigan/_utils.py b/src/so_vits_svc_fork/modules/decoders/hifigan/_utils.py
new file mode 100644
index 00000000..8f862c11
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/decoders/hifigan/_utils.py
@@ -0,0 +1,15 @@
+from logging import getLogger
+
+# matplotlib.use("Agg")
+
+LOG = getLogger(__name__)
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
diff --git a/src/so_vits_svc_fork/modules/decoders/mb_istft/__init__.py b/src/so_vits_svc_fork/modules/decoders/mb_istft/__init__.py
new file mode 100644
index 00000000..1ba61d27
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/decoders/mb_istft/__init__.py
@@ -0,0 +1,15 @@
+from ._generators import (
+    Multiband_iSTFT_Generator,
+    Multistream_iSTFT_Generator,
+    iSTFT_Generator,
+)
+from ._loss import subband_stft_loss
+from ._pqmf import PQMF
+
+__all__ = [
+    "subband_stft_loss",
+    "PQMF",
+    "iSTFT_Generator",
+    "Multiband_iSTFT_Generator",
+    "Multistream_iSTFT_Generator",
+]
diff --git a/src/so_vits_svc_fork/modules/decoders/mb_istft/_generators.py b/src/so_vits_svc_fork/modules/decoders/mb_istft/_generators.py
new file mode 100644
index 00000000..77f4b88a
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/decoders/mb_istft/_generators.py
@@ -0,0 +1,376 @@
+import math
+
+import torch
+from torch import nn
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, weight_norm
+
+from ....modules import modules
+from ....modules.commons import get_padding, init_weights
+from ._pqmf import PQMF
+from ._stft import TorchSTFT
+
+
+class iSTFT_Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gen_istft_n_fft,
+        gen_istft_hop_size,
+        gin_channels=0,
+    ):
+        super().__init__()
+        # self.h = h
+        self.gen_istft_n_fft = gen_istft_n_fft
+        self.gen_istft_hop_size = gen_istft_hop_size
+
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = weight_norm(
+            Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.post_n_fft = self.gen_istft_n_fft
+        self.conv_post = weight_norm(Conv1d(ch, self.post_n_fft + 2, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+        self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
+        self.stft = TorchSTFT(
+            filter_length=self.gen_istft_n_fft,
+            hop_length=self.gen_istft_hop_size,
+            win_length=self.gen_istft_n_fft,
+        )
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.reflection_pad(x)
+        x = self.conv_post(x)
+        spec = torch.exp(x[:, : self.post_n_fft // 2 + 1, :])
+        phase = math.pi * torch.sin(x[:, self.post_n_fft // 2 + 1 :, :])
+        out = self.stft.inverse(spec, phase).to(x.device)
+        return out, None
+
+    def remove_weight_norm(self):
+        print("Removing weight norm...")
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+
+class Multiband_iSTFT_Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gen_istft_n_fft,
+        gen_istft_hop_size,
+        subbands,
+        gin_channels=0,
+    ):
+        super().__init__()
+        # self.h = h
+        self.subbands = subbands
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = weight_norm(
+            Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.post_n_fft = gen_istft_n_fft
+        self.ups.apply(init_weights)
+        self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
+        self.reshape_pixelshuffle = []
+
+        self.subband_conv_post = weight_norm(
+            Conv1d(ch, self.subbands * (self.post_n_fft + 2), 7, 1, padding=3)
+        )
+
+        self.subband_conv_post.apply(init_weights)
+
+        self.gen_istft_n_fft = gen_istft_n_fft
+        self.gen_istft_hop_size = gen_istft_hop_size
+
+    def forward(self, x, g=None):
+        stft = TorchSTFT(
+            filter_length=self.gen_istft_n_fft,
+            hop_length=self.gen_istft_hop_size,
+            win_length=self.gen_istft_n_fft,
+        ).to(x.device)
+        pqmf = PQMF(x.device, subbands=self.subbands).to(x.device, dtype=x.dtype)
+
+        x = self.conv_pre(x)  # [B, ch, length]
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        x = F.leaky_relu(x)
+        x = self.reflection_pad(x)
+        x = self.subband_conv_post(x)
+        x = torch.reshape(
+            x, (x.shape[0], self.subbands, x.shape[1] // self.subbands, x.shape[-1])
+        )
+
+        spec = torch.exp(x[:, :, : self.post_n_fft // 2 + 1, :])
+        phase = math.pi * torch.sin(x[:, :, self.post_n_fft // 2 + 1 :, :])
+
+        y_mb_hat = stft.inverse(
+            torch.reshape(
+                spec,
+                (
+                    spec.shape[0] * self.subbands,
+                    self.gen_istft_n_fft // 2 + 1,
+                    spec.shape[-1],
+                ),
+            ),
+            torch.reshape(
+                phase,
+                (
+                    phase.shape[0] * self.subbands,
+                    self.gen_istft_n_fft // 2 + 1,
+                    phase.shape[-1],
+                ),
+            ),
+        )
+        y_mb_hat = torch.reshape(
+            y_mb_hat, (x.shape[0], self.subbands, 1, y_mb_hat.shape[-1])
+        )
+        y_mb_hat = y_mb_hat.squeeze(-2)
+
+        y_g_hat = pqmf.synthesis(y_mb_hat)
+
+        return y_g_hat, y_mb_hat
+
+    def remove_weight_norm(self):
+        print("Removing weight norm...")
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+class Multistream_iSTFT_Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gen_istft_n_fft,
+        gen_istft_hop_size,
+        subbands,
+        gin_channels=0,
+    ):
+        super().__init__()
+        # self.h = h
+        self.subbands = subbands
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = weight_norm(
+            Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.post_n_fft = gen_istft_n_fft
+        self.ups.apply(init_weights)
+        self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
+        self.reshape_pixelshuffle = []
+
+        self.subband_conv_post = weight_norm(
+            Conv1d(ch, self.subbands * (self.post_n_fft + 2), 7, 1, padding=3)
+        )
+
+        self.subband_conv_post.apply(init_weights)
+
+        self.gen_istft_n_fft = gen_istft_n_fft
+        self.gen_istft_hop_size = gen_istft_hop_size
+
+        updown_filter = torch.zeros(
+            (self.subbands, self.subbands, self.subbands)
+        ).float()
+        for k in range(self.subbands):
+            updown_filter[k, k, 0] = 1.0
+        self.register_buffer("updown_filter", updown_filter)
+        self.multistream_conv_post = weight_norm(
+            Conv1d(
+                self.subbands, 1, kernel_size=63, bias=False, padding=get_padding(63, 1)
+            )
+        )
+        self.multistream_conv_post.apply(init_weights)
+
+    def forward(self, x, g=None):
+        stft = TorchSTFT(
+            filter_length=self.gen_istft_n_fft,
+            hop_length=self.gen_istft_hop_size,
+            win_length=self.gen_istft_n_fft,
+        ).to(x.device)
+        # pqmf = PQMF(x.device)
+
+        x = self.conv_pre(x)  # [B, ch, length]
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        x = F.leaky_relu(x)
+        x = self.reflection_pad(x)
+        x = self.subband_conv_post(x)
+        x = torch.reshape(
+            x, (x.shape[0], self.subbands, x.shape[1] // self.subbands, x.shape[-1])
+        )
+
+        spec = torch.exp(x[:, :, : self.post_n_fft // 2 + 1, :])
+        phase = math.pi * torch.sin(x[:, :, self.post_n_fft // 2 + 1 :, :])
+
+        y_mb_hat = stft.inverse(
+            torch.reshape(
+                spec,
+                (
+                    spec.shape[0] * self.subbands,
+                    self.gen_istft_n_fft // 2 + 1,
+                    spec.shape[-1],
+                ),
+            ),
+            torch.reshape(
+                phase,
+                (
+                    phase.shape[0] * self.subbands,
+                    self.gen_istft_n_fft // 2 + 1,
+                    phase.shape[-1],
+                ),
+            ),
+        )
+        y_mb_hat = torch.reshape(
+            y_mb_hat, (x.shape[0], self.subbands, 1, y_mb_hat.shape[-1])
+        )
+        y_mb_hat = y_mb_hat.squeeze(-2)
+
+        y_mb_hat = F.conv_transpose1d(
+            y_mb_hat,
+            self.updown_filter.cuda(x.device) * self.subbands,
+            stride=self.subbands,
+        )
+
+        y_g_hat = self.multistream_conv_post(y_mb_hat)
+
+        return y_g_hat, y_mb_hat
+
+    def remove_weight_norm(self):
+        print("Removing weight norm...")
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
diff --git a/src/so_vits_svc_fork/modules/decoders/mb_istft/_loss.py b/src/so_vits_svc_fork/modules/decoders/mb_istft/_loss.py
new file mode 100644
index 00000000..7e8b7634
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/decoders/mb_istft/_loss.py
@@ -0,0 +1,11 @@
+from ._stft_loss import MultiResolutionSTFTLoss
+
+
+def subband_stft_loss(h, y_mb, y_hat_mb):
+    sub_stft_loss = MultiResolutionSTFTLoss(
+        h.train.fft_sizes, h.train.hop_sizes, h.train.win_lengths
+    )
+    y_mb = y_mb.view(-1, y_mb.size(2))
+    y_hat_mb = y_hat_mb.view(-1, y_hat_mb.size(2))
+    sub_sc_loss, sub_mag_loss = sub_stft_loss(y_hat_mb[:, : y_mb.size(-1)], y_mb)
+    return sub_sc_loss + sub_mag_loss
diff --git a/src/so_vits_svc_fork/modules/decoders/mb_istft/_pqmf.py b/src/so_vits_svc_fork/modules/decoders/mb_istft/_pqmf.py
new file mode 100644
index 00000000..981c1739
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/decoders/mb_istft/_pqmf.py
@@ -0,0 +1,130 @@
+# Copyright 2020 Tomoki Hayashi
+#  MIT License (https://opensource.org/licenses/MIT)
+
+"""Pseudo QMF modules."""
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from scipy.signal import kaiser
+
+
+def design_prototype_filter(taps=62, cutoff_ratio=0.15, beta=9.0):
+    """Design prototype filter for PQMF.
+    This method is based on `A Kaiser window approach for the design of prototype
+    filters of cosine modulated filterbanks`_.
+    Args:
+        taps (int): The number of filter taps.
+        cutoff_ratio (float): Cut-off frequency ratio.
+        beta (float): Beta coefficient for kaiser window.
+    Returns:
+        ndarray: Impluse response of prototype filter (taps + 1,).
+    .. _`A Kaiser window approach for the design of prototype filters of cosine modulated filterbanks`:
+        https://ieeexplore.ieee.org/abstract/document/681427
+    """
+    # check the arguments are valid
+    assert taps % 2 == 0, "The number of taps mush be even number."
+    assert 0.0 < cutoff_ratio < 1.0, "Cutoff ratio must be > 0.0 and < 1.0."
+
+    # make initial filter
+    omega_c = np.pi * cutoff_ratio
+    with np.errstate(invalid="ignore"):
+        h_i = np.sin(omega_c * (np.arange(taps + 1) - 0.5 * taps)) / (
+            np.pi * (np.arange(taps + 1) - 0.5 * taps)
+        )
+    h_i[taps // 2] = np.cos(0) * cutoff_ratio  # fix nan due to indeterminate form
+
+    # apply kaiser window
+    w = kaiser(taps + 1, beta)
+    h = h_i * w
+
+    return h
+
+
+class PQMF(torch.nn.Module):
+    """PQMF module.
+    This module is based on `Near-perfect-reconstruction pseudo-QMF banks`_.
+    .. _`Near-perfect-reconstruction pseudo-QMF banks`:
+        https://ieeexplore.ieee.org/document/258122
+    """
+
+    def __init__(self, device, subbands=8, taps=62, cutoff_ratio=0.15, beta=9.0):
+        """Initialize PQMF module.
+        Args:
+            subbands (int): The number of subbands.
+            taps (int): The number of filter taps.
+            cutoff_ratio (float): Cut-off frequency ratio.
+            beta (float): Beta coefficient for kaiser window.
+        """
+        super().__init__()
+
+        # define filter coefficient
+        h_proto = design_prototype_filter(taps, cutoff_ratio, beta)
+        h_analysis = np.zeros((subbands, len(h_proto)))
+        h_synthesis = np.zeros((subbands, len(h_proto)))
+        for k in range(subbands):
+            h_analysis[k] = (
+                2
+                * h_proto
+                * np.cos(
+                    (2 * k + 1)
+                    * (np.pi / (2 * subbands))
+                    * (np.arange(taps + 1) - ((taps - 1) / 2))
+                    + (-1) ** k * np.pi / 4
+                )
+            )
+            h_synthesis[k] = (
+                2
+                * h_proto
+                * np.cos(
+                    (2 * k + 1)
+                    * (np.pi / (2 * subbands))
+                    * (np.arange(taps + 1) - ((taps - 1) / 2))
+                    - (-1) ** k * np.pi / 4
+                )
+            )
+
+        # convert to tensor
+        analysis_filter = torch.from_numpy(h_analysis).float().unsqueeze(1).cuda(device)
+        synthesis_filter = (
+            torch.from_numpy(h_synthesis).float().unsqueeze(0).cuda(device)
+        )
+
+        # register coefficients as buffer
+        self.register_buffer("analysis_filter", analysis_filter)
+        self.register_buffer("synthesis_filter", synthesis_filter)
+
+        # filter for downsampling & upsampling
+        updown_filter = torch.zeros((subbands, subbands, subbands)).float().cuda(device)
+        for k in range(subbands):
+            updown_filter[k, k, 0] = 1.0
+        self.register_buffer("updown_filter", updown_filter)
+        self.subbands = subbands
+
+        # keep padding info
+        self.pad_fn = torch.nn.ConstantPad1d(taps // 2, 0.0)
+
+    def analysis(self, x):
+        """Analysis with PQMF.
+        Args:
+            x (Tensor): Input tensor (B, 1, T).
+        Returns:
+            Tensor: Output tensor (B, subbands, T // subbands).
+        """
+        x = F.conv1d(self.pad_fn(x), self.analysis_filter)
+        return F.conv1d(x, self.updown_filter, stride=self.subbands)
+
+    def synthesis(self, x):
+        """Synthesis with PQMF.
+        Args:
+            x (Tensor): Input tensor (B, subbands, T // subbands).
+        Returns:
+            Tensor: Output tensor (B, 1, T).
+        """
+        # NOTE(kan-bayashi): Power will be dreased so here multiply by # subbands.
+        #   Not sure this is the correct way, it is better to check again.
+        # TODO(kan-bayashi): Understand the reconstruction procedure
+        x = F.conv_transpose1d(
+            x, self.updown_filter * self.subbands, stride=self.subbands
+        )
+        return F.conv1d(self.pad_fn(x), self.synthesis_filter)
diff --git a/src/so_vits_svc_fork/modules/decoders/mb_istft/_stft.py b/src/so_vits_svc_fork/modules/decoders/mb_istft/_stft.py
new file mode 100644
index 00000000..8a111dca
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/decoders/mb_istft/_stft.py
@@ -0,0 +1,248 @@
+"""
+BSD 3-Clause License
+Copyright (c) 2017, Prem Seetharaman
+All rights reserved.
+* Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are met:
+* Redistributions of source code must retain the above copyright notice,
+  this list of conditions and the following disclaimer.
+* Redistributions in binary form must reproduce the above copyright notice, this
+  list of conditions and the following disclaimer in the
+  documentation and/or other materials provided with the distribution.
+* Neither the name of the copyright holder nor the names of its
+  contributors may be used to endorse or promote products derived from this
+  software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+"""
+
+import librosa.util as librosa_util
+import numpy as np
+import torch
+import torch.nn.functional as F
+from librosa.util import pad_center, tiny
+from scipy.signal import get_window
+from torch.autograd import Variable
+
+
+def window_sumsquare(
+    window,
+    n_frames,
+    hop_length=200,
+    win_length=800,
+    n_fft=800,
+    dtype=np.float32,
+    norm=None,
+):
+    """
+    # from librosa 0.6
+    Compute the sum-square envelope of a window function at a given hop length.
+    This is used to estimate modulation effects induced by windowing
+    observations in short-time fourier transforms.
+    Parameters
+    ----------
+    window : string, tuple, number, callable, or list-like
+        Window specification, as in `get_window`
+    n_frames : int > 0
+        The number of analysis frames
+    hop_length : int > 0
+        The number of samples to advance between frames
+    win_length : [optional]
+        The length of the window function.  By default, this matches `n_fft`.
+    n_fft : int > 0
+        The length of each analysis frame.
+    dtype : np.dtype
+        The data type of the output
+    Returns
+    -------
+    wss : np.ndarray, shape=`(n_fft + hop_length * (n_frames - 1))`
+        The sum-squared envelope of the window function
+    """
+    if win_length is None:
+        win_length = n_fft
+
+    n = n_fft + hop_length * (n_frames - 1)
+    x = np.zeros(n, dtype=dtype)
+
+    # Compute the squared window at the desired length
+    win_sq = get_window(window, win_length, fftbins=True)
+    win_sq = librosa_util.normalize(win_sq, norm=norm) ** 2
+    win_sq = librosa_util.pad_center(win_sq, n_fft)
+
+    # Fill the envelope
+    for i in range(n_frames):
+        sample = i * hop_length
+        x[sample : min(n, sample + n_fft)] += win_sq[: max(0, min(n_fft, n - sample))]
+    return x
+
+
+class STFT(torch.nn.Module):
+    """adapted from Prem Seetharaman's https://github.com/pseeth/pytorch-stft"""
+
+    def __init__(
+        self, filter_length=800, hop_length=200, win_length=800, window="hann"
+    ):
+        super().__init__()
+        self.filter_length = filter_length
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.window = window
+        self.forward_transform = None
+        scale = self.filter_length / self.hop_length
+        fourier_basis = np.fft.fft(np.eye(self.filter_length))
+
+        cutoff = int(self.filter_length / 2 + 1)
+        fourier_basis = np.vstack(
+            [np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])]
+        )
+
+        forward_basis = torch.FloatTensor(fourier_basis[:, None, :])
+        inverse_basis = torch.FloatTensor(
+            np.linalg.pinv(scale * fourier_basis).T[:, None, :]
+        )
+
+        if window is not None:
+            assert filter_length >= win_length
+            # get window and zero center pad it to filter_length
+            fft_window = get_window(window, win_length, fftbins=True)
+            fft_window = pad_center(fft_window, filter_length)
+            fft_window = torch.from_numpy(fft_window).float()
+
+            # window the bases
+            forward_basis *= fft_window
+            inverse_basis *= fft_window
+
+        self.register_buffer("forward_basis", forward_basis.float())
+        self.register_buffer("inverse_basis", inverse_basis.float())
+
+    def transform(self, input_data):
+        num_batches = input_data.size(0)
+        num_samples = input_data.size(1)
+
+        self.num_samples = num_samples
+
+        # similar to librosa, reflect-pad the input
+        input_data = input_data.view(num_batches, 1, num_samples)
+        input_data = F.pad(
+            input_data.unsqueeze(1),
+            (int(self.filter_length / 2), int(self.filter_length / 2), 0, 0),
+            mode="reflect",
+        )
+        input_data = input_data.squeeze(1)
+
+        forward_transform = F.conv1d(
+            input_data,
+            Variable(self.forward_basis, requires_grad=False),
+            stride=self.hop_length,
+            padding=0,
+        )
+
+        cutoff = int((self.filter_length / 2) + 1)
+        real_part = forward_transform[:, :cutoff, :]
+        imag_part = forward_transform[:, cutoff:, :]
+
+        magnitude = torch.sqrt(real_part**2 + imag_part**2)
+        phase = torch.autograd.Variable(torch.atan2(imag_part.data, real_part.data))
+
+        return magnitude, phase
+
+    def inverse(self, magnitude, phase):
+        recombine_magnitude_phase = torch.cat(
+            [magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
+        )
+
+        inverse_transform = F.conv_transpose1d(
+            recombine_magnitude_phase,
+            Variable(self.inverse_basis, requires_grad=False),
+            stride=self.hop_length,
+            padding=0,
+        )
+
+        if self.window is not None:
+            window_sum = window_sumsquare(
+                self.window,
+                magnitude.size(-1),
+                hop_length=self.hop_length,
+                win_length=self.win_length,
+                n_fft=self.filter_length,
+                dtype=np.float32,
+            )
+            # remove modulation effects
+            approx_nonzero_indices = torch.from_numpy(
+                np.where(window_sum > tiny(window_sum))[0]
+            )
+            window_sum = torch.autograd.Variable(
+                torch.from_numpy(window_sum), requires_grad=False
+            )
+            window_sum = (
+                window_sum.to(inverse_transform.device())
+                if magnitude.is_cuda
+                else window_sum
+            )
+            inverse_transform[:, :, approx_nonzero_indices] /= window_sum[
+                approx_nonzero_indices
+            ]
+
+            # scale by hop ratio
+            inverse_transform *= float(self.filter_length) / self.hop_length
+
+        inverse_transform = inverse_transform[:, :, int(self.filter_length / 2) :]
+        inverse_transform = inverse_transform[:, :, : -int(self.filter_length / 2) :]
+
+        return inverse_transform
+
+    def forward(self, input_data):
+        self.magnitude, self.phase = self.transform(input_data)
+        reconstruction = self.inverse(self.magnitude, self.phase)
+        return reconstruction
+
+
+class TorchSTFT(torch.nn.Module):
+    def __init__(
+        self, filter_length=800, hop_length=200, win_length=800, window="hann"
+    ):
+        super().__init__()
+        self.filter_length = filter_length
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.window = torch.from_numpy(
+            get_window(window, win_length, fftbins=True).astype(np.float32)
+        )
+
+    def transform(self, input_data):
+        forward_transform = torch.stft(
+            input_data,
+            self.filter_length,
+            self.hop_length,
+            self.win_length,
+            window=self.window,
+            return_complex=True,
+        )
+
+        return torch.abs(forward_transform), torch.angle(forward_transform)
+
+    def inverse(self, magnitude, phase):
+        inverse_transform = torch.istft(
+            magnitude * torch.exp(phase * 1j),
+            self.filter_length,
+            self.hop_length,
+            self.win_length,
+            window=self.window.to(magnitude.device),
+        )
+
+        return inverse_transform.unsqueeze(
+            -2
+        )  # unsqueeze to stay consistent with conv_transpose1d implementation
+
+    def forward(self, input_data):
+        self.magnitude, self.phase = self.transform(input_data)
+        reconstruction = self.inverse(self.magnitude, self.phase)
+        return reconstruction
diff --git a/src/so_vits_svc_fork/modules/decoders/mb_istft/_stft_loss.py b/src/so_vits_svc_fork/modules/decoders/mb_istft/_stft_loss.py
new file mode 100644
index 00000000..c685cb02
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/decoders/mb_istft/_stft_loss.py
@@ -0,0 +1,142 @@
+# Copyright 2019 Tomoki Hayashi
+#  MIT License (https://opensource.org/licenses/MIT)
+
+"""STFT-based Loss modules."""
+
+import torch
+import torch.nn.functional as F
+
+
+def stft(x, fft_size, hop_size, win_length, window):
+    """Perform STFT and convert to magnitude spectrogram.
+    Args:
+        x (Tensor): Input signal tensor (B, T).
+        fft_size (int): FFT size.
+        hop_size (int): Hop size.
+        win_length (int): Window length.
+        window (str): Window function type.
+    Returns:
+        Tensor: Magnitude spectrogram (B, #frames, fft_size // 2 + 1).
+    """
+    x_stft = torch.stft(
+        x, fft_size, hop_size, win_length, window.to(x.device), return_complex=False
+    )
+    real = x_stft[..., 0]
+    imag = x_stft[..., 1]
+
+    # NOTE(kan-bayashi): clamp is needed to avoid nan or inf
+    return torch.sqrt(torch.clamp(real**2 + imag**2, min=1e-7)).transpose(2, 1)
+
+
+class SpectralConvergengeLoss(torch.nn.Module):
+    """Spectral convergence loss module."""
+
+    def __init__(self):
+        """Initialize spectral convergence loss module."""
+        super().__init__()
+
+    def forward(self, x_mag, y_mag):
+        """Calculate forward propagation.
+        Args:
+            x_mag (Tensor): Magnitude spectrogram of predicted signal (B, #frames, #freq_bins).
+            y_mag (Tensor): Magnitude spectrogram of groundtruth signal (B, #frames, #freq_bins).
+        Returns:
+            Tensor: Spectral convergence loss value.
+        """
+        return torch.norm(y_mag - x_mag) / torch.norm(
+            y_mag
+        )  # MB-iSTFT-VITS changed here due to codespell
+
+
+class LogSTFTMagnitudeLoss(torch.nn.Module):
+    """Log STFT magnitude loss module."""
+
+    def __init__(self):
+        """Initialize los STFT magnitude loss module."""
+        super().__init__()
+
+    def forward(self, x_mag, y_mag):
+        """Calculate forward propagation.
+        Args:
+            x_mag (Tensor): Magnitude spectrogram of predicted signal (B, #frames, #freq_bins).
+            y_mag (Tensor): Magnitude spectrogram of groundtruth signal (B, #frames, #freq_bins).
+        Returns:
+            Tensor: Log STFT magnitude loss value.
+        """
+        return F.l1_loss(torch.log(y_mag), torch.log(x_mag))
+
+
+class STFTLoss(torch.nn.Module):
+    """STFT loss module."""
+
+    def __init__(
+        self, fft_size=1024, shift_size=120, win_length=600, window="hann_window"
+    ):
+        """Initialize STFT loss module."""
+        super().__init__()
+        self.fft_size = fft_size
+        self.shift_size = shift_size
+        self.win_length = win_length
+        self.window = getattr(torch, window)(win_length)
+        self.spectral_convergenge_loss = SpectralConvergengeLoss()
+        self.log_stft_magnitude_loss = LogSTFTMagnitudeLoss()
+
+    def forward(self, x, y):
+        """Calculate forward propagation.
+        Args:
+            x (Tensor): Predicted signal (B, T).
+            y (Tensor): Groundtruth signal (B, T).
+        Returns:
+            Tensor: Spectral convergence loss value.
+            Tensor: Log STFT magnitude loss value.
+        """
+        x_mag = stft(x, self.fft_size, self.shift_size, self.win_length, self.window)
+        y_mag = stft(y, self.fft_size, self.shift_size, self.win_length, self.window)
+        sc_loss = self.spectral_convergenge_loss(x_mag, y_mag)
+        mag_loss = self.log_stft_magnitude_loss(x_mag, y_mag)
+
+        return sc_loss, mag_loss
+
+
+class MultiResolutionSTFTLoss(torch.nn.Module):
+    """Multi resolution STFT loss module."""
+
+    def __init__(
+        self,
+        fft_sizes=[1024, 2048, 512],
+        hop_sizes=[120, 240, 50],
+        win_lengths=[600, 1200, 240],
+        window="hann_window",
+    ):
+        """Initialize Multi resolution STFT loss module.
+        Args:
+            fft_sizes (list): List of FFT sizes.
+            hop_sizes (list): List of hop sizes.
+            win_lengths (list): List of window lengths.
+            window (str): Window function type.
+        """
+        super().__init__()
+        assert len(fft_sizes) == len(hop_sizes) == len(win_lengths)
+        self.stft_losses = torch.nn.ModuleList()
+        for fs, ss, wl in zip(fft_sizes, hop_sizes, win_lengths):
+            self.stft_losses += [STFTLoss(fs, ss, wl, window)]
+
+    def forward(self, x, y):
+        """Calculate forward propagation.
+        Args:
+            x (Tensor): Predicted signal (B, T).
+            y (Tensor): Groundtruth signal (B, T).
+        Returns:
+            Tensor: Multi resolution spectral convergence loss value.
+            Tensor: Multi resolution log STFT magnitude loss value.
+        """
+        sc_loss = 0.0
+        mag_loss = 0.0
+        for f in self.stft_losses:
+            sc_l, mag_l = f(x, y)
+            sc_loss += sc_l
+            mag_loss += mag_l
+        sc_loss /= len(self.stft_losses)
+        mag_loss /= len(self.stft_losses)
+
+        return sc_loss, mag_loss
diff --git a/src/so_vits_svc_fork/modules/descriminators.py b/src/so_vits_svc_fork/modules/descriminators.py
new file mode 100644
index 00000000..a59b1e5f
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/descriminators.py
@@ -0,0 +1,177 @@
+import torch
+from torch import nn
+from torch.nn import AvgPool1d, Conv1d, Conv2d
+from torch.nn import functional as F
+from torch.nn.utils import spectral_norm, weight_norm
+
+from so_vits_svc_fork.modules import modules as modules
+from so_vits_svc_fork.modules.commons import get_padding
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super().__init__()
+        self.period = period
+        self.use_spectral_norm = use_spectral_norm
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    Conv2d(
+                        1,
+                        32,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        1024,
+                        1024,
+                        (kernel_size, 1),
+                        1,
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+            ]
+        )
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super().__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(Conv1d(1, 16, 15, 1, padding=7)),
+                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
+                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+            ]
+        )
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super().__init__()
+        periods = [2, 3, 5, 7, 11]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class MultiScaleDiscriminator(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.discriminators = nn.ModuleList(
+            [
+                DiscriminatorS(use_spectral_norm=True),
+                DiscriminatorS(),
+                DiscriminatorS(),
+            ]
+        )
+        self.meanpools = nn.ModuleList(
+            [AvgPool1d(4, 2, padding=2), AvgPool1d(4, 2, padding=2)]
+        )
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            if i != 0:
+                y = self.meanpools[i - 1](y)
+                y_hat = self.meanpools[i - 1](y_hat)
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
diff --git a/src/so_vits_svc_fork/modules/encoders.py b/src/so_vits_svc_fork/modules/encoders.py
new file mode 100644
index 00000000..4894aa5c
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/encoders.py
@@ -0,0 +1,136 @@
+import torch
+from torch import nn
+
+from so_vits_svc_fork.modules import attentions as attentions
+from so_vits_svc_fork.modules import commons as commons
+from so_vits_svc_fork.modules import modules as modules
+
+
+class SpeakerEncoder(torch.nn.Module):
+    def __init__(
+        self,
+        mel_n_channels=80,
+        model_num_layers=3,
+        model_hidden_size=256,
+        model_embedding_size=256,
+    ):
+        super().__init__()
+        self.lstm = nn.LSTM(
+            mel_n_channels, model_hidden_size, model_num_layers, batch_first=True
+        )
+        self.linear = nn.Linear(model_hidden_size, model_embedding_size)
+        self.relu = nn.ReLU()
+
+    def forward(self, mels):
+        self.lstm.flatten_parameters()
+        _, (hidden, _) = self.lstm(mels)
+        embeds_raw = self.relu(self.linear(hidden[-1]))
+        return embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True)
+
+    def compute_partial_slices(self, total_frames, partial_frames, partial_hop):
+        mel_slices = []
+        for i in range(0, total_frames - partial_frames, partial_hop):
+            mel_range = torch.arange(i, i + partial_frames)
+            mel_slices.append(mel_range)
+
+        return mel_slices
+
+    def embed_utterance(self, mel, partial_frames=128, partial_hop=64):
+        mel_len = mel.size(1)
+        last_mel = mel[:, -partial_frames:]
+
+        if mel_len > partial_frames:
+            mel_slices = self.compute_partial_slices(
+                mel_len, partial_frames, partial_hop
+            )
+            mels = list(mel[:, s] for s in mel_slices)
+            mels.append(last_mel)
+            mels = torch.stack(tuple(mels), 0).squeeze(1)
+
+            with torch.no_grad():
+                partial_embeds = self(mels)
+            embed = torch.mean(partial_embeds, axis=0).unsqueeze(0)
+            # embed = embed / torch.linalg.norm(embed, 2)
+        else:
+            with torch.no_grad():
+                embed = self(last_mel)
+
+        return embed
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None):
+        # print(x.shape,x_lengths.shape)
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+
+class TextEncoder(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        n_layers,
+        gin_channels=0,
+        filter_channels=None,
+        n_heads=None,
+        p_dropout=None,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+        self.f0_emb = nn.Embedding(256, hidden_channels)
+
+        self.enc_ = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+
+    def forward(self, x, x_mask, f0=None, noice_scale=1):
+        x = x + self.f0_emb(f0).transpose(1, 2)
+        x = self.enc_(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs) * noice_scale) * x_mask
+
+        return z, m, logs, x_mask
diff --git a/src/so_vits_svc_fork/modules/flows.py b/src/so_vits_svc_fork/modules/flows.py
new file mode 100644
index 00000000..9abcba21
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/flows.py
@@ -0,0 +1,48 @@
+from torch import nn
+
+from so_vits_svc_fork.modules import modules as modules
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
diff --git a/src/so_vits_svc_fork/modules/mel_processing.py b/src/so_vits_svc_fork/modules/mel_processing.py
index ff3ad9cd..0de6e54b 100644
--- a/src/so_vits_svc_fork/modules/mel_processing.py
+++ b/src/so_vits_svc_fork/modules/mel_processing.py
@@ -99,9 +99,9 @@ def mel_spectrogram_torch(
     y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False
 ):
     if torch.min(y) < -1.0:
-        LOG.info("min value is ", torch.min(y))
+        LOG.info(f"min value is {torch.min(y)}")
     if torch.max(y) > 1.0:
-        LOG.info("max value is ", torch.max(y))
+        LOG.info(f"max value is {torch.max(y)}")
 
     global mel_basis, hann_window
     dtype_device = str(y.dtype) + "_" + str(y.device)
diff --git a/src/so_vits_svc_fork/modules/synthesizers.py b/src/so_vits_svc_fork/modules/synthesizers.py
new file mode 100644
index 00000000..0f796215
--- /dev/null
+++ b/src/so_vits_svc_fork/modules/synthesizers.py
@@ -0,0 +1,233 @@
+import warnings
+from logging import getLogger
+from typing import Any, Literal, Sequence
+
+import torch
+from torch import nn
+
+import so_vits_svc_fork.f0
+from so_vits_svc_fork.f0 import f0_to_coarse
+from so_vits_svc_fork.modules import commons as commons
+from so_vits_svc_fork.modules.decoders.f0 import F0Decoder
+from so_vits_svc_fork.modules.decoders.hifigan import Generator
+from so_vits_svc_fork.modules.decoders.mb_istft import (
+    Multiband_iSTFT_Generator,
+    Multistream_iSTFT_Generator,
+    iSTFT_Generator,
+)
+from so_vits_svc_fork.modules.encoders import Encoder, TextEncoder
+from so_vits_svc_fork.modules.flows import ResidualCouplingBlock
+
+LOG = getLogger(__name__)
+
+
+class SynthesizerTrn(nn.Module):
+    """
+    Synthesizer for Training
+    """
+
+    def __init__(
+        self,
+        spec_channels: int,
+        segment_size: int,
+        inter_channels: int,
+        hidden_channels: int,
+        filter_channels: int,
+        n_heads: int,
+        n_layers: int,
+        kernel_size: int,
+        p_dropout: int,
+        resblock: str,
+        resblock_kernel_sizes: Sequence[int],
+        resblock_dilation_sizes: Sequence[Sequence[int]],
+        upsample_rates: Sequence[int],
+        upsample_initial_channel: int,
+        upsample_kernel_sizes: Sequence[int],
+        gin_channels: int,
+        ssl_dim: int,
+        n_speakers: int,
+        sampling_rate: int = 44100,
+        type_: Literal["hifi-gan", "istft", "ms-istft", "mb-istft"] = "hifi-gan",
+        gen_istft_n_fft: int = 16,
+        gen_istft_hop_size: int = 4,
+        subbands: int = 4,
+        **kwargs: Any,
+    ):
+        super().__init__()
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        self.ssl_dim = ssl_dim
+        self.n_speakers = n_speakers
+        self.sampling_rate = sampling_rate
+        self.type_ = type_
+        self.gen_istft_n_fft = gen_istft_n_fft
+        self.gen_istft_hop_size = gen_istft_hop_size
+        self.subbands = subbands
+        if kwargs:
+            warnings.warn(f"Unused arguments: {kwargs}")
+
+        self.emb_g = nn.Embedding(n_speakers, gin_channels)
+
+        if ssl_dim is None:
+            self.pre = nn.LazyConv1d(hidden_channels, kernel_size=5, padding=2)
+        else:
+            self.pre = nn.Conv1d(ssl_dim, hidden_channels, kernel_size=5, padding=2)
+
+        self.enc_p = TextEncoder(
+            inter_channels,
+            hidden_channels,
+            filter_channels=filter_channels,
+            n_heads=n_heads,
+            n_layers=n_layers,
+            kernel_size=kernel_size,
+            p_dropout=p_dropout,
+        )
+
+        LOG.info(f"Decoder type: {type_}")
+        if type_ == "hifi-gan":
+            hps = {
+                "sampling_rate": sampling_rate,
+                "inter_channels": inter_channels,
+                "resblock": resblock,
+                "resblock_kernel_sizes": resblock_kernel_sizes,
+                "resblock_dilation_sizes": resblock_dilation_sizes,
+                "upsample_rates": upsample_rates,
+                "upsample_initial_channel": upsample_initial_channel,
+                "upsample_kernel_sizes": upsample_kernel_sizes,
+                "gin_channels": gin_channels,
+            }
+            self.dec = Generator(h=hps)
+            self.mb = False
+        else:
+            hps = {
+                "initial_channel": inter_channels,
+                "resblock": resblock,
+                "resblock_kernel_sizes": resblock_kernel_sizes,
+                "resblock_dilation_sizes": resblock_dilation_sizes,
+                "upsample_rates": upsample_rates,
+                "upsample_initial_channel": upsample_initial_channel,
+                "upsample_kernel_sizes": upsample_kernel_sizes,
+                "gin_channels": gin_channels,
+                "gen_istft_n_fft": gen_istft_n_fft,
+                "gen_istft_hop_size": gen_istft_hop_size,
+                "subbands": subbands,
+            }
+
+            # gen_istft_n_fft, gen_istft_hop_size, subbands
+            if type_ == "istft":
+                del hps["subbands"]
+                self.dec = iSTFT_Generator(**hps)
+            elif type_ == "ms-istft":
+                self.dec = Multistream_iSTFT_Generator(**hps)
+            elif type_ == "mb-istft":
+                self.dec = Multiband_iSTFT_Generator(**hps)
+            else:
+                raise ValueError(f"Unknown type: {type_}")
+            self.mb = True
+
+        self.enc_q = Encoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels
+        )
+        self.f0_decoder = F0Decoder(
+            1,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            spk_channels=gin_channels,
+        )
+        self.emb_uv = nn.Embedding(2, hidden_channels)
+
+    def forward(self, c, f0, uv, spec, g=None, c_lengths=None, spec_lengths=None):
+        g = self.emb_g(g).transpose(1, 2)
+        # ssl prenet
+        x_mask = torch.unsqueeze(commons.sequence_mask(c_lengths, c.size(2)), 1).to(
+            c.dtype
+        )
+        x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1, 2)
+
+        # f0 predict
+        lf0 = 2595.0 * torch.log10(1.0 + f0.unsqueeze(1) / 700.0) / 500
+        norm_lf0 = so_vits_svc_fork.f0.normalize_f0(lf0, x_mask, uv)
+        pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
+
+        # encoder
+        z_ptemp, m_p, logs_p, _ = self.enc_p(x, x_mask, f0=f0_to_coarse(f0))
+        z, m_q, logs_q, spec_mask = self.enc_q(spec, spec_lengths, g=g)
+
+        # flow
+        z_p = self.flow(z, spec_mask, g=g)
+        z_slice, pitch_slice, ids_slice = commons.rand_slice_segments_with_pitch(
+            z, f0, spec_lengths, self.segment_size
+        )
+
+        # MB-iSTFT-VITS
+        if self.mb:
+            o, o_mb = self.dec(z_slice, g=g)
+        # HiFi-GAN
+        else:
+            o = self.dec(z_slice, g=g, f0=pitch_slice)
+            o_mb = None
+        return (
+            o,
+            o_mb,
+            ids_slice,
+            spec_mask,
+            (z, z_p, m_p, logs_p, m_q, logs_q),
+            pred_lf0,
+            norm_lf0,
+            lf0,
+        )
+
+    def infer(self, c, f0, uv, g=None, noice_scale=0.35, predict_f0=False):
+        c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device)
+        g = self.emb_g(g).transpose(1, 2)
+        x_mask = torch.unsqueeze(commons.sequence_mask(c_lengths, c.size(2)), 1).to(
+            c.dtype
+        )
+        x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1, 2)
+
+        if predict_f0:
+            lf0 = 2595.0 * torch.log10(1.0 + f0.unsqueeze(1) / 700.0) / 500
+            norm_lf0 = so_vits_svc_fork.f0.normalize_f0(
+                lf0, x_mask, uv, random_scale=False
+            )
+            pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
+            f0 = (700 * (torch.pow(10, pred_lf0 * 500 / 2595) - 1)).squeeze(1)
+
+        z_p, m_p, logs_p, c_mask = self.enc_p(
+            x, x_mask, f0=f0_to_coarse(f0), noice_scale=noice_scale
+        )
+        z = self.flow(z_p, c_mask, g=g, reverse=True)
+
+        # MB-iSTFT-VITS
+        if self.mb:
+            o, o_mb = self.dec(z * c_mask, g=g)
+        else:
+            o = self.dec(z * c_mask, g=g, f0=f0)
+        return o
diff --git a/src/so_vits_svc_fork/onnx_export.py b/src/so_vits_svc_fork/onnx_export.py
deleted file mode 100644
index 401c266d..00000000
--- a/src/so_vits_svc_fork/onnx_export.py
+++ /dev/null
@@ -1,65 +0,0 @@
-from __future__ import annotations
-
-from pathlib import Path
-
-import torch
-
-from . import utils
-from .onnxexport.model_onnx import SynthesizerTrn
-
-
-def onnx_export(
-    input_path: Path | str,
-    output_path: Path | str,
-    config_path: Path | str,
-    device: str | torch.device = "cpu",
-):
-    input_path = Path(input_path)
-    output_path = Path(output_path)
-    config_path = Path(config_path)
-    hps = utils.get_hparams_from_file(config_path.as_posix())
-    SVCVITS = SynthesizerTrn(
-        hps.data.filter_length // 2 + 1,
-        hps.train.segment_size // hps.data.hop_length,
-        **hps.model,
-    )
-    _ = utils.load_checkpoint(input_path.as_posix(), SVCVITS, None)
-    _ = SVCVITS.eval().to(device)
-    for i in SVCVITS.parameters():
-        i.requires_grad = False
-
-    test_hidden_unit = torch.rand(1, 10, 256)
-    test_pitch = torch.rand(1, 10)
-    test_mel2ph = torch.LongTensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]).unsqueeze(0)
-    test_uv = torch.ones(1, 10, dtype=torch.float32)
-    test_noise = torch.randn(1, 192, 10)
-    test_sid = torch.LongTensor([0])
-    input_names = ["c", "f0", "mel2ph", "uv", "noise", "sid"]
-    output_names = [
-        "audio",
-    ]
-
-    torch.onnx.export(
-        SVCVITS,
-        (
-            test_hidden_unit.to(device),
-            test_pitch.to(device),
-            test_mel2ph.to(device),
-            test_uv.to(device),
-            test_noise.to(device),
-            test_sid.to(device),
-        ),
-        output_path.as_posix(),
-        dynamic_axes={
-            "c": [0, 1],
-            "f0": [1],
-            "mel2ph": [1],
-            "uv": [1],
-            "noise": [2],
-        },
-        do_constant_folding=False,
-        opset_version=16,
-        verbose=False,
-        input_names=input_names,
-        output_names=output_names,
-    )
diff --git a/src/so_vits_svc_fork/onnxexport/model_onnx.py b/src/so_vits_svc_fork/onnxexport/model_onnx.py
deleted file mode 100644
index d450b9c6..00000000
--- a/src/so_vits_svc_fork/onnxexport/model_onnx.py
+++ /dev/null
@@ -1,404 +0,0 @@
-import torch
-from torch import nn
-from torch.nn import Conv1d, Conv2d
-from torch.nn import functional as F
-from torch.nn.utils import spectral_norm, weight_norm
-
-from .. import utils
-from ..modules import attentions, commons, modules
-from ..modules.commons import get_padding
-from ..utils import f0_to_coarse
-from ..vdecoder.hifigan.models import Generator
-
-
-class ResidualCouplingBlock(nn.Module):
-    def __init__(
-        self,
-        channels,
-        hidden_channels,
-        kernel_size,
-        dilation_rate,
-        n_layers,
-        n_flows=4,
-        gin_channels=0,
-    ):
-        super().__init__()
-        self.channels = channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.dilation_rate = dilation_rate
-        self.n_layers = n_layers
-        self.n_flows = n_flows
-        self.gin_channels = gin_channels
-
-        self.flows = nn.ModuleList()
-        for i in range(n_flows):
-            self.flows.append(
-                modules.ResidualCouplingLayer(
-                    channels,
-                    hidden_channels,
-                    kernel_size,
-                    dilation_rate,
-                    n_layers,
-                    gin_channels=gin_channels,
-                    mean_only=True,
-                )
-            )
-            self.flows.append(modules.Flip())
-
-    def forward(self, x, x_mask, g=None, reverse=False):
-        if not reverse:
-            for flow in self.flows:
-                x, _ = flow(x, x_mask, g=g, reverse=reverse)
-        else:
-            for flow in reversed(self.flows):
-                x = flow(x, x_mask, g=g, reverse=reverse)
-        return x
-
-
-class Encoder(nn.Module):
-    def __init__(
-        self,
-        in_channels,
-        out_channels,
-        hidden_channels,
-        kernel_size,
-        dilation_rate,
-        n_layers,
-        gin_channels=0,
-    ):
-        super().__init__()
-        self.in_channels = in_channels
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.dilation_rate = dilation_rate
-        self.n_layers = n_layers
-        self.gin_channels = gin_channels
-
-        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
-        self.enc = modules.WN(
-            hidden_channels,
-            kernel_size,
-            dilation_rate,
-            n_layers,
-            gin_channels=gin_channels,
-        )
-        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
-
-    def forward(self, x, x_lengths, g=None):
-        # print(x.shape,x_lengths.shape)
-        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
-            x.dtype
-        )
-        x = self.pre(x) * x_mask
-        x = self.enc(x, x_mask, g=g)
-        stats = self.proj(x) * x_mask
-        m, logs = torch.split(stats, self.out_channels, dim=1)
-        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
-        return z, m, logs, x_mask
-
-
-class TextEncoder(nn.Module):
-    def __init__(
-        self,
-        out_channels,
-        hidden_channels,
-        kernel_size,
-        n_layers,
-        gin_channels=0,
-        filter_channels=None,
-        n_heads=None,
-        p_dropout=None,
-    ):
-        super().__init__()
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.n_layers = n_layers
-        self.gin_channels = gin_channels
-        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
-        self.f0_emb = nn.Embedding(256, hidden_channels)
-
-        self.enc_ = attentions.Encoder(
-            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
-        )
-
-    def forward(self, x, x_mask, f0=None, z=None):
-        x = x + self.f0_emb(f0).transpose(1, 2)
-        x = self.enc_(x * x_mask, x_mask)
-        stats = self.proj(x) * x_mask
-        m, logs = torch.split(stats, self.out_channels, dim=1)
-        z = (m + z * torch.exp(logs)) * x_mask
-        return z, m, logs, x_mask
-
-
-class DiscriminatorP(torch.nn.Module):
-    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
-        super().__init__()
-        self.period = period
-        self.use_spectral_norm = use_spectral_norm
-        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
-        self.convs = nn.ModuleList(
-            [
-                norm_f(
-                    Conv2d(
-                        1,
-                        32,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        32,
-                        128,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        128,
-                        512,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        512,
-                        1024,
-                        (kernel_size, 1),
-                        (stride, 1),
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-                norm_f(
-                    Conv2d(
-                        1024,
-                        1024,
-                        (kernel_size, 1),
-                        1,
-                        padding=(get_padding(kernel_size, 1), 0),
-                    )
-                ),
-            ]
-        )
-        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
-
-    def forward(self, x):
-        fmap = []
-
-        # 1d to 2d
-        b, c, t = x.shape
-        if t % self.period != 0:  # pad first
-            n_pad = self.period - (t % self.period)
-            x = F.pad(x, (0, n_pad), "reflect")
-            t = t + n_pad
-        x = x.view(b, c, t // self.period, self.period)
-
-        for l in self.convs:
-            x = l(x)
-            x = F.leaky_relu(x, modules.LRELU_SLOPE)
-            fmap.append(x)
-        x = self.conv_post(x)
-        fmap.append(x)
-        x = torch.flatten(x, 1, -1)
-
-        return x, fmap
-
-
-class DiscriminatorS(torch.nn.Module):
-    def __init__(self, use_spectral_norm=False):
-        super().__init__()
-        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
-        self.convs = nn.ModuleList(
-            [
-                norm_f(Conv1d(1, 16, 15, 1, padding=7)),
-                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
-                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
-                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
-                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
-                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
-            ]
-        )
-        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
-
-    def forward(self, x):
-        fmap = []
-
-        for l in self.convs:
-            x = l(x)
-            x = F.leaky_relu(x, modules.LRELU_SLOPE)
-            fmap.append(x)
-        x = self.conv_post(x)
-        fmap.append(x)
-        x = torch.flatten(x, 1, -1)
-
-        return x, fmap
-
-
-class F0Decoder(nn.Module):
-    def __init__(
-        self,
-        out_channels,
-        hidden_channels,
-        filter_channels,
-        n_heads,
-        n_layers,
-        kernel_size,
-        p_dropout,
-        spk_channels=0,
-    ):
-        super().__init__()
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.filter_channels = filter_channels
-        self.n_heads = n_heads
-        self.n_layers = n_layers
-        self.kernel_size = kernel_size
-        self.p_dropout = p_dropout
-        self.spk_channels = spk_channels
-
-        self.prenet = nn.Conv1d(hidden_channels, hidden_channels, 3, padding=1)
-        self.decoder = attentions.FFT(
-            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
-        )
-        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
-        self.f0_prenet = nn.Conv1d(1, hidden_channels, 3, padding=1)
-        self.cond = nn.Conv1d(spk_channels, hidden_channels, 1)
-
-    def forward(self, x, norm_f0, x_mask, spk_emb=None):
-        x = torch.detach(x)
-        if spk_emb is not None:
-            x = x + self.cond(spk_emb)
-        x += self.f0_prenet(norm_f0)
-        x = self.prenet(x) * x_mask
-        x = self.decoder(x * x_mask, x_mask)
-        x = self.proj(x) * x_mask
-        return x
-
-
-class SynthesizerTrn(nn.Module):
-    """
-    Synthesizer for Training
-    """
-
-    def __init__(
-        self,
-        spec_channels,
-        segment_size,
-        inter_channels,
-        hidden_channels,
-        filter_channels,
-        n_heads,
-        n_layers,
-        kernel_size,
-        p_dropout,
-        resblock,
-        resblock_kernel_sizes,
-        resblock_dilation_sizes,
-        upsample_rates,
-        upsample_initial_channel,
-        upsample_kernel_sizes,
-        gin_channels,
-        ssl_dim,
-        n_speakers,
-        sampling_rate=44100,
-        **kwargs
-    ):
-        super().__init__()
-        self.spec_channels = spec_channels
-        self.inter_channels = inter_channels
-        self.hidden_channels = hidden_channels
-        self.filter_channels = filter_channels
-        self.n_heads = n_heads
-        self.n_layers = n_layers
-        self.kernel_size = kernel_size
-        self.p_dropout = p_dropout
-        self.resblock = resblock
-        self.resblock_kernel_sizes = resblock_kernel_sizes
-        self.resblock_dilation_sizes = resblock_dilation_sizes
-        self.upsample_rates = upsample_rates
-        self.upsample_initial_channel = upsample_initial_channel
-        self.upsample_kernel_sizes = upsample_kernel_sizes
-        self.segment_size = segment_size
-        self.gin_channels = gin_channels
-        self.ssl_dim = ssl_dim
-        self.emb_g = nn.Embedding(n_speakers, gin_channels)
-
-        self.pre = nn.Conv1d(ssl_dim, hidden_channels, kernel_size=5, padding=2)
-
-        self.enc_p = TextEncoder(
-            inter_channels,
-            hidden_channels,
-            filter_channels=filter_channels,
-            n_heads=n_heads,
-            n_layers=n_layers,
-            kernel_size=kernel_size,
-            p_dropout=p_dropout,
-        )
-        hps = {
-            "sampling_rate": sampling_rate,
-            "inter_channels": inter_channels,
-            "resblock": resblock,
-            "resblock_kernel_sizes": resblock_kernel_sizes,
-            "resblock_dilation_sizes": resblock_dilation_sizes,
-            "upsample_rates": upsample_rates,
-            "upsample_initial_channel": upsample_initial_channel,
-            "upsample_kernel_sizes": upsample_kernel_sizes,
-            "gin_channels": gin_channels,
-        }
-        self.dec = Generator(h=hps)
-        self.enc_q = Encoder(
-            spec_channels,
-            inter_channels,
-            hidden_channels,
-            5,
-            1,
-            16,
-            gin_channels=gin_channels,
-        )
-        self.flow = ResidualCouplingBlock(
-            inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels
-        )
-        self.f0_decoder = F0Decoder(
-            1,
-            hidden_channels,
-            filter_channels,
-            n_heads,
-            n_layers,
-            kernel_size,
-            p_dropout,
-            spk_channels=gin_channels,
-        )
-        self.emb_uv = nn.Embedding(2, hidden_channels)
-        self.predict_f0 = False
-
-    def forward(self, c, f0, mel2ph, uv, noise=None, g=None):
-        decoder_inp = F.pad(c, [0, 0, 1, 0])
-        mel2ph_ = mel2ph.unsqueeze(2).repeat([1, 1, c.shape[-1]])
-        c = torch.gather(decoder_inp, 1, mel2ph_).transpose(1, 2)  # [B, T, H]
-
-        c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device)
-        g = g.unsqueeze(0)
-        g = self.emb_g(g).transpose(1, 2)
-        x_mask = torch.unsqueeze(commons.sequence_mask(c_lengths, c.size(2)), 1).to(
-            c.dtype
-        )
-        x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1, 2)
-
-        if self.predict_f0:
-            lf0 = 2595.0 * torch.log10(1.0 + f0.unsqueeze(1) / 700.0) / 500
-            norm_lf0 = utils.normalize_f0(lf0, x_mask, uv, random_scale=False)
-            pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
-            f0 = (700 * (torch.pow(10, pred_lf0 * 500 / 2595) - 1)).squeeze(1)
-
-        z_p, m_p, logs_p, c_mask = self.enc_p(x, x_mask, f0=f0_to_coarse(f0), z=noise)
-        z = self.flow(z_p, c_mask, g=g, reverse=True)
-        o = self.dec(z * c_mask, g=g, f0=f0)
-        return o
diff --git a/src/so_vits_svc_fork/onnxexport/__init__.py b/src/so_vits_svc_fork/preprocessing/__init__.py
similarity index 100%
rename from src/so_vits_svc_fork/onnxexport/__init__.py
rename to src/so_vits_svc_fork/preprocessing/__init__.py
diff --git a/src/so_vits_svc_fork/preprocessing/config_templates/quickvc.json b/src/so_vits_svc_fork/preprocessing/config_templates/quickvc.json
new file mode 100644
index 00000000..dbef0a69
--- /dev/null
+++ b/src/so_vits_svc_fork/preprocessing/config_templates/quickvc.json
@@ -0,0 +1,69 @@
+{
+  "train": {
+    "log_interval": 100,
+    "eval_interval": 200,
+    "seed": 1234,
+    "epochs": 10000,
+    "learning_rate": 0.0001,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 18,
+    "fp16_run": false,
+    "lr_decay": 0.999875,
+    "segment_size": 10240,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0,
+    "use_sr": true,
+    "max_speclen": 512,
+    "port": "8001",
+    "keep_ckpts": 3,
+    "fft_sizes": [768, 1366, 342],
+    "hop_sizes": [60, 120, 20],
+    "win_lengths": [300, 600, 120],
+    "window": "hann_window"
+  },
+  "data": {
+    "training_files": "filelists/44k/train.txt",
+    "validation_files": "filelists/44k/val.txt",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 44100,
+    "filter_length": 2048,
+    "hop_length": 512,
+    "win_length": 2048,
+    "n_mel_channels": 80,
+    "mel_fmin": 0.0,
+    "mel_fmax": 22050,
+    "contentvec_final_proj": false
+  },
+  "model": {
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3, 7, 11],
+    "resblock_dilation_sizes": [
+      [1, 3, 5],
+      [1, 3, 5],
+      [1, 3, 5]
+    ],
+    "upsample_rates": [8, 4],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [32, 16],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 256,
+    "ssl_dim": 768,
+    "n_speakers": 200,
+    "type_": "ms-istft",
+    "gen_istft_n_fft": 16,
+    "gen_istft_hop_size": 4,
+    "subbands": 4
+  },
+  "spk": {}
+}
diff --git a/src/so_vits_svc_fork/configs_template/config_template.json b/src/so_vits_svc_fork/preprocessing/config_templates/so-vits-svc-4.0v1-legacy.json
similarity index 100%
rename from src/so_vits_svc_fork/configs_template/config_template.json
rename to src/so_vits_svc_fork/preprocessing/config_templates/so-vits-svc-4.0v1-legacy.json
diff --git a/src/so_vits_svc_fork/preprocessing/config_templates/so-vits-svc-4.0v1.json b/src/so_vits_svc_fork/preprocessing/config_templates/so-vits-svc-4.0v1.json
new file mode 100644
index 00000000..d4c8d46f
--- /dev/null
+++ b/src/so_vits_svc_fork/preprocessing/config_templates/so-vits-svc-4.0v1.json
@@ -0,0 +1,62 @@
+{
+  "train": {
+    "log_interval": 100,
+    "eval_interval": 200,
+    "seed": 1234,
+    "epochs": 10000,
+    "learning_rate": 0.0001,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 18,
+    "fp16_run": false,
+    "lr_decay": 0.999875,
+    "segment_size": 10240,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0,
+    "use_sr": true,
+    "max_speclen": 512,
+    "port": "8001",
+    "keep_ckpts": 3
+  },
+  "data": {
+    "training_files": "filelists/44k/train.txt",
+    "validation_files": "filelists/44k/val.txt",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 44100,
+    "filter_length": 2048,
+    "hop_length": 512,
+    "win_length": 2048,
+    "n_mel_channels": 80,
+    "mel_fmin": 0.0,
+    "mel_fmax": 22050,
+    "contentvec_final_proj": false
+  },
+  "model": {
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3, 7, 11],
+    "resblock_dilation_sizes": [
+      [1, 3, 5],
+      [1, 3, 5],
+      [1, 3, 5]
+    ],
+    "upsample_rates": [8, 8, 2, 2, 2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16, 16, 4, 4, 4],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 256,
+    "ssl_dim": 768,
+    "n_speakers": 200,
+    "type_": "hifi-gan"
+  },
+  "spk": {}
+}
diff --git a/src/so_vits_svc_fork/preprocess_flist_config.py b/src/so_vits_svc_fork/preprocessing/preprocess_flist_config.py
similarity index 92%
rename from src/so_vits_svc_fork/preprocess_flist_config.py
rename to src/so_vits_svc_fork/preprocessing/preprocess_flist_config.py
index 5db387b0..a6642bee 100644
--- a/src/so_vits_svc_fork/preprocess_flist_config.py
+++ b/src/so_vits_svc_fork/preprocessing/preprocess_flist_config.py
@@ -11,6 +11,7 @@
 from tqdm import tqdm
 
 LOG = getLogger(__name__)
+CONFIG_TEMPLATE_DIR = Path(__file__).parent / "config_templates"
 
 
 def preprocess_config(
@@ -19,6 +20,7 @@ def preprocess_config(
     val_list_path: Path | str,
     test_list_path: Path | str,
     config_path: Path | str,
+    config_name: str,
 ):
     input_dir = Path(input_dir)
     train_list_path = Path(train_list_path)
@@ -76,7 +78,9 @@ def preprocess_config(
     config = deepcopy(
         json.loads(
             (
-                Path(__file__).parent / "configs_template" / "config_template.json"
+                CONFIG_TEMPLATE_DIR / f"{config_name}.json"
+                if not config_name.endswith(".json")
+                else config_name
             ).read_text()
         )
     )
diff --git a/src/so_vits_svc_fork/preprocess_hubert_f0.py b/src/so_vits_svc_fork/preprocessing/preprocess_hubert_f0.py
similarity index 60%
rename from src/so_vits_svc_fork/preprocess_hubert_f0.py
rename to src/so_vits_svc_fork/preprocessing/preprocess_hubert_f0.py
index d4f383d0..3c7618a5 100644
--- a/src/so_vits_svc_fork/preprocess_hubert_f0.py
+++ b/src/so_vits_svc_fork/preprocessing/preprocess_hubert_f0.py
@@ -8,17 +8,22 @@
 import librosa
 import numpy as np
 import torch
-from joblib import Parallel, cpu_count, delayed
+from joblib import Parallel, delayed
 from tqdm import tqdm
 
-from . import utils
+import so_vits_svc_fork.f0
+from so_vits_svc_fork import utils
+
+from ..utils import get_total_gpu_memory
 from .preprocess_utils import check_hubert_min_duration
-from .utils import HUBERT_SAMPLING_RATE
 
 LOG = getLogger(__name__)
+HUBERT_MEMORY = 1600
+HUBERT_MEMORY_CREPE = 2600
 
 
 def _process_one(
+    *,
     filepath: Path,
     hubert_model,
     sampling_rate: int,
@@ -26,6 +31,7 @@ def _process_one(
     device: Literal["cuda", "cpu"] = "cuda",
     f0_method: Literal["crepe", "crepe-tiny", "parselmouth", "dio", "harvest"] = "dio",
     force_rebuild: bool = False,
+    legacy_final_proj: bool = False,
 ):
     audio, sr = librosa.load(filepath, sr=sampling_rate)
 
@@ -36,11 +42,9 @@ def _process_one(
     # Compute HuBERT content
     soft_path = filepath.parent / (filepath.name + ".soft.pt")
     if (not soft_path.exists()) or force_rebuild:
-        wav16k = librosa.resample(
-            audio, orig_sr=sampling_rate, target_sr=HUBERT_SAMPLING_RATE
+        c = utils.get_content(
+            hubert_model, audio, device, sr=sr, legacy_final_proj=legacy_final_proj
         )
-        wav16k = torch.from_numpy(wav16k).to(device)
-        c = utils.get_hubert_content(hubert_model, wav_16k_tensor=wav16k)
         torch.save(c.cpu(), soft_path)
     else:
         LOG.info(f"Skip {filepath} because {soft_path} exists.")
@@ -48,7 +52,7 @@ def _process_one(
     # Compute f0
     f0_path = filepath.parent / (filepath.name + ".f0.npy")
     if (not f0_path.exists()) or force_rebuild:
-        f0 = utils.compute_f0(
+        f0 = so_vits_svc_fork.f0.compute_f0(
             audio, sampling_rate=sampling_rate, hop_length=hop_length, method=f0_method
         )
         np.save(f0_path, f0)
@@ -58,49 +62,62 @@ def _process_one(
 
 
 def _process_batch(
+    *,
     filepaths: Iterable[Path],
     sampling_rate: int,
     hop_length: int,
     pbar_position: int,
     f0_method: Literal["crepe", "crepe-tiny", "parselmouth", "dio", "harvest"] = "dio",
     force_rebuild: bool = False,
+    legacy_final_proj: bool = False,
 ):
     device = "cuda" if torch.cuda.is_available() else "cpu"
-    hubert_model = utils.get_hubert_model().to(device)
+    hubert_model = utils.get_hubert_model(device)
 
     for filepath in tqdm(filepaths, position=pbar_position):
         _process_one(
-            filepath,
-            hubert_model,
-            sampling_rate,
-            hop_length,
-            device,
-            f0_method,
-            force_rebuild,
+            filepath=filepath,
+            hubert_model=hubert_model,
+            sampling_rate=sampling_rate,
+            hop_length=hop_length,
+            device=device,
+            f0_method=f0_method,
+            force_rebuild=force_rebuild,
+            legacy_final_proj=legacy_final_proj,
         )
 
 
 def preprocess_hubert_f0(
     input_dir: Path | str,
     config_path: Path | str,
-    n_jobs: int = 4,
+    n_jobs: int | None = None,
     f0_method: Literal["crepe", "crepe-tiny", "parselmouth", "dio", "harvest"] = "dio",
     force_rebuild: bool = False,
 ):
     input_dir = Path(input_dir)
     config_path = Path(config_path)
-    utils.ensure_hubert_model()
-    hps = utils.get_hparams_from_file(config_path)
-    sampling_rate = hps.data.sampling_rate
-    hop_length = hps.data.hop_length
+    utils.ensure_pretrained_model(".", "contentvec")
+    hps = utils.get_hparams(config_path)
+    if n_jobs is None:
+        memory = get_total_gpu_memory("free")
+        n_jobs = memory // (
+            HUBERT_MEMORY_CREPE if f0_method == "crepe" else HUBERT_MEMORY
+        )
+        LOG.info(f"n_jobs automatically set to {n_jobs}, memory: {memory} MiB")
 
     filepaths = list(input_dir.rglob("*.wav"))
-    n_jobs = min(cpu_count(), len(filepaths) // 32 + 1, n_jobs)
+    n_jobs = min(len(filepaths) // 16 + 1, n_jobs)
     shuffle(filepaths)
     filepath_chunks = np.array_split(filepaths, n_jobs)
     Parallel(n_jobs=n_jobs)(
         delayed(_process_batch)(
-            chunk, sampling_rate, hop_length, pbar_position, f0_method, force_rebuild
+            filepaths=chunk,
+            sampling_rate=hps.data.sampling_rate,
+            hop_length=hps.data.hop_length,
+            pbar_position=pbar_position,
+            f0_method=f0_method,
+            force_rebuild=force_rebuild,
+            legacy_final_proj=hps.data.__dict__.get("contentvec_final_proj", True),
         )
         for (pbar_position, chunk) in enumerate(filepath_chunks)
     )
diff --git a/src/so_vits_svc_fork/preprocess_resample.py b/src/so_vits_svc_fork/preprocessing/preprocess_resample.py
similarity index 100%
rename from src/so_vits_svc_fork/preprocess_resample.py
rename to src/so_vits_svc_fork/preprocessing/preprocess_resample.py
diff --git a/src/so_vits_svc_fork/preprocess_speaker_diarization.py b/src/so_vits_svc_fork/preprocessing/preprocess_speaker_diarization.py
similarity index 100%
rename from src/so_vits_svc_fork/preprocess_speaker_diarization.py
rename to src/so_vits_svc_fork/preprocessing/preprocess_speaker_diarization.py
diff --git a/src/so_vits_svc_fork/preprocess_split.py b/src/so_vits_svc_fork/preprocessing/preprocess_split.py
similarity index 100%
rename from src/so_vits_svc_fork/preprocess_split.py
rename to src/so_vits_svc_fork/preprocessing/preprocess_split.py
diff --git a/src/so_vits_svc_fork/preprocess_utils.py b/src/so_vits_svc_fork/preprocessing/preprocess_utils.py
similarity index 100%
rename from src/so_vits_svc_fork/preprocess_utils.py
rename to src/so_vits_svc_fork/preprocessing/preprocess_utils.py
diff --git a/src/so_vits_svc_fork/spec_gen.py b/src/so_vits_svc_fork/spec_gen.py
deleted file mode 100644
index b0ba544b..00000000
--- a/src/so_vits_svc_fork/spec_gen.py
+++ /dev/null
@@ -1,23 +0,0 @@
-import json
-
-from tqdm import tqdm
-
-from .data_utils import TextAudioSpeakerLoader
-from .utils import HParams
-
-config_path = "configs/config.json"
-with open(config_path) as f:
-    data = f.read()
-config = json.loads(data)
-hps = HParams(**config)
-
-train_dataset = TextAudioSpeakerLoader("filelists/train.txt", hps)
-test_dataset = TextAudioSpeakerLoader("filelists/test.txt", hps)
-eval_dataset = TextAudioSpeakerLoader("filelists/val.txt", hps)
-
-for _ in tqdm(train_dataset):
-    pass
-for _ in tqdm(eval_dataset):
-    pass
-for _ in tqdm(test_dataset):
-    pass
diff --git a/src/so_vits_svc_fork/train.py b/src/so_vits_svc_fork/train.py
index ec09859c..c8272882 100644
--- a/src/so_vits_svc_fork/train.py
+++ b/src/so_vits_svc_fork/train.py
@@ -9,6 +9,7 @@
 import torch
 import torch.distributed as dist
 import torch.multiprocessing as mp
+from torch import nn
 from torch.cuda.amp import GradScaler, autocast
 from torch.nn import functional as F
 from torch.nn.parallel import DistributedDataParallel as DDP
@@ -16,16 +17,17 @@
 from torch.utils.tensorboard import SummaryWriter
 from tqdm import trange
 
+import so_vits_svc_fork.f0
 import so_vits_svc_fork.modules.commons as commons
+import so_vits_svc_fork.utils
 
 from . import utils
 from .data_utils import TextAudioCollate, TextAudioSpeakerLoader
-from .models import MultiPeriodDiscriminator, SynthesizerTrn
+from .hparams import HParams
+from .modules.descriminators import MultiPeriodDiscriminator
 from .modules.losses import discriminator_loss, feature_loss, generator_loss, kl_loss
 from .modules.mel_processing import mel_spectrogram_torch, spec_to_mel_torch
-
-# os.environ['TORCH_DISTRIBUTED_DEBUG'] = 'INFO'
-
+from .modules.synthesizers import SynthesizerTrn
 
 LOG = getLogger(__name__)
 torch.backends.cudnn.benchmark = True
@@ -33,30 +35,29 @@
 start_time = time.time()
 
 
-def train(config_path: Path | str, model_path: Path | str):
+def train(
+    config_path: Path | str, model_path: Path | str, reset_optimizer: bool = False
+):
     """Assume Single Node Multi GPUs Training Only"""
     config_path = Path(config_path)
     model_path = Path(model_path)
     if not torch.cuda.is_available():
         raise RuntimeError("CUDA is not available.")
-    utils.ensure_pretrained_model(model_path)
-    hps = utils.get_hparams(config_path, model_path)
 
+    hps = utils.get_backup_hparams(config_path, model_path)
+    utils.ensure_pretrained_model(model_path, hps.model.get("type_", "hifi-gan"))
     n_gpus = torch.cuda.device_count()
     os.environ["MASTER_ADDR"] = "localhost"
     os.environ["MASTER_PORT"] = hps.train.port
 
     mp.spawn(
-        run,
+        _run,
         nprocs=n_gpus,
-        args=(
-            n_gpus,
-            hps,
-        ),
+        args=(n_gpus, hps, reset_optimizer),
     )
 
 
-def run(rank, n_gpus, hps):
+def _run(rank: int, n_gpus: int, hps: HParams, reset_optimizer: bool = False):
     global global_step
     if rank == 0:
         LOG.info(hps)
@@ -114,31 +115,34 @@ def run(rank, n_gpus, hps):
         betas=hps.train.betas,
         eps=hps.train.eps,
     )
-    net_g = DDP(net_g, device_ids=[rank])  # , find_unused_parameters=True)
+    net_g = DDP(net_g, device_ids=[rank])
     net_d = DDP(net_d, device_ids=[rank])
 
-    skip_optimizer = False
-    try:
-        _, _, _, epoch_str = utils.load_checkpoint(
-            utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"),
-            net_g,
-            optim_g,
-            skip_optimizer,
-        )
-        _, _, _, epoch_str = utils.load_checkpoint(
-            utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"),
-            net_d,
-            optim_d,
-            skip_optimizer,
-        )
-        epoch_str = max(epoch_str, 1)
-        global_step = (epoch_str - 1) * len(train_loader)
-    except Exception as e:
-        LOG.exception(e)
-        LOG.info("No checkpoint found, start from scratch")
+    latest_g_path = utils.latest_checkpoint_path(hps.model_dir, "G_*.pth")
+    latest_d_path = utils.latest_checkpoint_path(hps.model_dir, "D_*.pth")
+    if latest_g_path is not None and latest_d_path is not None:
+        try:
+            _, _, _, epoch_str = utils.load_checkpoint(
+                latest_g_path,
+                net_g,
+                optim_g,
+                reset_optimizer,
+            )
+            _, _, _, epoch_str = utils.load_checkpoint(
+                latest_d_path,
+                net_d,
+                optim_d,
+                reset_optimizer,
+            )
+            epoch_str = max(epoch_str, 1)
+            global_step = (epoch_str - 1) * len(train_loader)
+        except Exception as e:
+            raise RuntimeError("Failed to load checkpoint") from e
+    else:
+        LOG.warning("No checkpoint found. Start from scratch.")
         epoch_str = 1
         global_step = 0
-    if skip_optimizer:
+    if reset_optimizer:
         epoch_str = 1
         global_step = 0
 
@@ -151,39 +155,54 @@ def run(rank, n_gpus, hps):
 
     scaler = GradScaler(enabled=hps.train.fp16_run)
 
-    LOG.info("Start training")
+    LOG.info(
+        "Start training..."
+        "Note: You do not need to wait until the progress bar is full."
+    )
 
-    for epoch in trange(epoch_str, hps.train.epochs + 1):
+    for epoch in trange(
+        epoch_str, hps.train.epochs + 1, initial=epoch_str, total=hps.train.epochs
+    ):
         if rank == 0:
-            train_and_evaluate(
+            _train_and_evaluate(
                 rank,
                 epoch,
                 hps,
-                [net_g, net_d],
-                [optim_g, optim_d],
-                [scheduler_g, scheduler_d],
+                (net_g, net_d),
+                (optim_g, optim_d),
+                (scheduler_g, scheduler_d),
                 scaler,
-                [train_loader, eval_loader],
-                [writer, writer_eval],
+                (train_loader, eval_loader),
+                (writer, writer_eval),
             )
         else:
-            train_and_evaluate(
+            _train_and_evaluate(
                 rank,
                 epoch,
                 hps,
-                [net_g, net_d],
-                [optim_g, optim_d],
-                [scheduler_g, scheduler_d],
+                (net_g, net_d),
+                (optim_g, optim_d),
+                (scheduler_g, scheduler_d),
                 scaler,
-                [train_loader, None],
+                (train_loader, None),
                 None,
             )
         scheduler_g.step()
         scheduler_d.step()
 
 
-def train_and_evaluate(
-    rank, epoch, hps, nets, optims, schedulers, scaler, loaders, writers
+def _train_and_evaluate(
+    rank: int,
+    epoch: int,
+    hps: HParams,
+    nets: tuple[nn.Module, nn.Module],
+    optims: tuple[torch.optim.Optimizer, torch.optim.Optimizer],
+    schedulers: tuple[
+        torch.optim.lr_scheduler.ExponentialLR, torch.optim.lr_scheduler.ExponentialLR
+    ],
+    scaler: GradScaler,
+    loaders: tuple[DataLoader, DataLoader | None],
+    writers: None | tuple[SummaryWriter, SummaryWriter],
 ):
     net_g, net_d = nets
     optim_g, optim_d = optims
@@ -217,6 +236,7 @@ def train_and_evaluate(
         with autocast(enabled=hps.train.fp16_run):
             (
                 y_hat,
+                y_hat_mb,
                 ids_slice,
                 z_mask,
                 (z, z_p, m_p, logs_p, m_q, logs_q),
@@ -267,6 +287,16 @@ def train_and_evaluate(
                 loss_gen, losses_gen = generator_loss(y_d_hat_g)
                 loss_lf0 = F.mse_loss(pred_lf0, lf0)
                 loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl + loss_lf0
+
+                # MB-iSTFT-VITS
+                loss_subband = torch.tensor(0.0)
+                if hps.model.get("type_") == "mb-istft":
+                    from .modules.decoders.mb_istft import PQMF, subband_stft_loss
+
+                    y_mb = PQMF(y.device, hps.model.subbands).analysis(y)
+                    loss_subband = subband_stft_loss(hps, y_mb, y_hat_mb)
+                loss_gen_all += loss_subband
+
         optim_g.zero_grad()
         scaler.scale(loss_gen_all).backward()
         scaler.unscale_(optim_g)
@@ -277,15 +307,21 @@ def train_and_evaluate(
         if rank == 0:
             if global_step % hps.train.log_interval == 0:
                 lr = optim_g.param_groups[0]["lr"]
-                losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_kl]
+                losses = {
+                    "discriminator": loss_disc.item(),
+                    "generator": loss_gen.item(),
+                    "feature_matching": loss_fm.item(),
+                    "melspectrogram": loss_mel.item(),
+                    "kl_divergence": loss_kl.item(),
+                }
+                if hps.model.get("type_") == "mb-istft":
+                    losses["subband_stft"] = loss_subband.item()
                 LOG.info(
                     "Train Epoch: {} [{:.0f}%]".format(
                         epoch, 100.0 * batch_idx / len(train_loader)
                     )
                 )
-                LOG.info(
-                    f"Losses: {[x.item() for x in losses]}, step: {global_step}, lr: {lr}"
-                )
+                LOG.info(f"Losses: {losses}, step: {global_step}, lr: {lr}")
 
                 scalar_dict = {
                     "loss/g/total": loss_gen_all,
@@ -302,6 +338,8 @@ def train_and_evaluate(
                         "loss/g/lf0": loss_lf0,
                     }
                 )
+                if hps.model.get("type_") == "mb-istft":
+                    scalar_dict["loss/g/subband"] = loss_subband
 
                 # scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)})
                 # scalar_dict.update({"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)})
@@ -316,11 +354,11 @@ def train_and_evaluate(
                     "all/mel": utils.plot_spectrogram_to_numpy(
                         mel[0].data.cpu().numpy()
                     ),
-                    "all/lf0": utils.plot_data_to_numpy(
+                    "all/lf0": so_vits_svc_fork.utils.plot_data_to_numpy(
                         lf0[0, 0, :].cpu().numpy(),
                         pred_lf0[0, 0, :].detach().cpu().numpy(),
                     ),
-                    "all/norm_lf0": utils.plot_data_to_numpy(
+                    "all/norm_lf0": so_vits_svc_fork.utils.plot_data_to_numpy(
                         lf0[0, 0, :].cpu().numpy(),
                         norm_lf0[0, 0, :].detach().cpu().numpy(),
                     ),
@@ -335,7 +373,7 @@ def train_and_evaluate(
 
             if global_step % hps.train.eval_interval == 0:
                 LOG.info("Saving checkpoints...")
-                evaluate(hps, net_g, eval_loader, writer_eval)
+                _evaluate(hps, net_g, eval_loader, writer_eval)
                 utils.save_checkpoint(
                     net_g,
                     optim_g,
@@ -368,7 +406,12 @@ def train_and_evaluate(
         start_time = now
 
 
-def evaluate(hps, generator, eval_loader, writer_eval):
+def _evaluate(
+    hps: HParams,
+    generator: torch.nn.Module,
+    eval_loader: torch.utils.data.DataLoader,
+    writer_eval: SummaryWriter,
+) -> None:
     generator.eval()
     image_dict = {}
     audio_dict = {}
diff --git a/src/so_vits_svc_fork/utils.py b/src/so_vits_svc_fork/utils.py
index 30e30fb6..c84ffbb8 100644
--- a/src/so_vits_svc_fork/utils.py
+++ b/src/so_vits_svc_fork/utils.py
@@ -2,393 +2,175 @@
 
 import json
 import re
+import subprocess
 from itertools import groupby
 from logging import getLogger
 from pathlib import Path
-from typing import Any, Literal
+from typing import Any, Literal, Sequence
 
+import matplotlib
+import matplotlib.pylab as plt
 import numpy as np
 import requests
 import torch
-import torchcrepe
 from cm_time import timer
-from numpy import dtype, float32, ndarray
+from fairseq import checkpoint_utils
+from fairseq.models.hubert.hubert import HubertModel
+from numpy import ndarray
 from scipy.io.wavfile import read
-from torch import FloatTensor, Tensor
 from tqdm import tqdm
 
+from so_vits_svc_fork.hparams import HParams
+
 LOG = getLogger(__name__)
-MATPLOTLIB_FLAG = False
-f0_bin = 256
-f0_max = 1100.0
-f0_min = 50.0
-f0_mel_min = 1127 * np.log(1 + f0_min / 700)
-f0_mel_max = 1127 * np.log(1 + f0_max / 700)
 HUBERT_SAMPLING_RATE = 16000
 
 
-# def normalize_f0(f0, random_scale=True):
-#     f0_norm = f0.clone()  # create a copy of the input Tensor
-#     batch_size, _, frame_length = f0_norm.shape
-#     for i in range(batch_size):
-#         means = torch.mean(f0_norm[i, 0, :])
-#         if random_scale:
-#             factor = random.uniform(0.8, 1.2)
-#         else:
-#             factor = 1
-#         f0_norm[i, 0, :] = (f0_norm[i, 0, :] - means) * factor
-#     return f0_norm
-# def normalize_f0(f0, random_scale=True):
-#     means = torch.mean(f0[:, 0, :], dim=1, keepdim=True)
-#     if random_scale:
-#         factor = torch.Tensor(f0.shape[0],1).uniform_(0.8, 1.2).to(f0.device)
-#     else:
-#         factor = torch.ones(f0.shape[0], 1, 1).to(f0.device)
-#     f0_norm = (f0 - means.unsqueeze(-1)) * factor.unsqueeze(-1)
-#     return f0_norm
-def normalize_f0(
-    f0: FloatTensor, x_mask: FloatTensor, uv: FloatTensor, random_scale=True
-) -> FloatTensor:
-    # calculate means based on x_mask
-    uv_sum = torch.sum(uv, dim=1, keepdim=True)
-    uv_sum[uv_sum == 0] = 9999
-    means = torch.sum(f0[:, 0, :] * uv, dim=1, keepdim=True) / uv_sum
-
-    if random_scale:
-        factor = torch.Tensor(f0.shape[0], 1).uniform_(0.8, 1.2).to(f0.device)
-    else:
-        factor = torch.ones(f0.shape[0], 1).to(f0.device)
-    # normalize f0 based on means and factor
-    f0_norm = (f0 - means.unsqueeze(-1)) * factor.unsqueeze(-1)
-    if torch.isnan(f0_norm).any():
-        exit(0)
-    return f0_norm * x_mask
-
-
-def plot_data_to_numpy(x: ndarray, y: ndarray) -> ndarray:
-    global MATPLOTLIB_FLAG
-    if not MATPLOTLIB_FLAG:
-        import matplotlib
-
-        matplotlib.use("Agg")
-        MATPLOTLIB_FLAG = True
-    import matplotlib.pylab as plt
-    import numpy as np
-
-    fig, ax = plt.subplots(figsize=(10, 2))
-    plt.plot(x)
-    plt.plot(y)
-    plt.tight_layout()
-
-    fig.canvas.draw()
-    data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
-    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
-    plt.close()
-    return data
-
-
-def interpolate_f0(
-    f0: ndarray[Any, dtype[float32]]
-) -> tuple[ndarray[Any, dtype[float32]], ndarray[Any, dtype[float32]]]:
-    data = np.reshape(f0, (f0.size, 1))
-
-    vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
-    vuv_vector[data > 0.0] = 1.0
-    vuv_vector[data <= 0.0] = 0.0
-
-    ip_data = data
-
-    frame_number = data.size
-    last_value = 0.0
-    for i in range(frame_number):
-        if data[i] <= 0.0:
-            j = i + 1
-            for j in range(i + 1, frame_number):
-                if data[j] > 0.0:
-                    break
-            if j < frame_number - 1:
-                if last_value > 0.0:
-                    step = (data[j] - data[i - 1]) / float(j - i)
-                    for k in range(i, j):
-                        ip_data[k] = data[i - 1] + step * (k - i + 1)
-                else:
-                    for k in range(i, j):
-                        ip_data[k] = data[j]
-            else:
-                for k in range(i, frame_number):
-                    ip_data[k] = last_value
-        else:
-            ip_data[i] = data[i]
-            last_value = data[i]
-
-    return ip_data[:, 0], vuv_vector[:, 0]
-
-
-def compute_f0_parselmouth(
-    wav_numpy: ndarray[Any, dtype[float32]],
-    p_len: None | int = None,
-    sampling_rate: int = 44100,
-    hop_length: int = 512,
-):
-    import parselmouth
-
-    x = wav_numpy
-    if p_len is None:
-        p_len = x.shape[0] // hop_length
-    else:
-        assert abs(p_len - x.shape[0] // hop_length) < 4, "pad length error"
-    time_step = hop_length / sampling_rate * 1000
-    f0_min = 50
-    f0_max = 1100
-    f0 = (
-        parselmouth.Sound(x, sampling_rate)
-        .to_pitch_ac(
-            time_step=time_step / 1000,
-            voicing_threshold=0.6,
-            pitch_floor=f0_min,
-            pitch_ceiling=f0_max,
-        )
-        .selected_array["frequency"]
-    )
-
-    pad_size = (p_len - len(f0) + 1) // 2
-    if pad_size > 0 or p_len - len(f0) - pad_size > 0:
-        f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
-    return f0
-
-
-def _resize_f0(
-    x: ndarray[Any, dtype[float32]], target_len: int
-) -> ndarray[Any, dtype[float32]]:
-    source = np.array(x)
-    source[source < 0.001] = np.nan
-    target = np.interp(
-        np.arange(0, len(source) * target_len, len(source)) / target_len,
-        np.arange(0, len(source)),
-        source,
-    )
-    res = np.nan_to_num(target)
-    return res
-
-
-def compute_f0_pyworld(
-    wav_numpy: ndarray[Any, dtype[float32]],
-    p_len: None | int = None,
-    sampling_rate: int = 44100,
-    hop_length: int = 512,
-    type_: Literal["dio", "harvest"] = "dio",
-):
-    import pyworld
-
-    if p_len is None:
-        p_len = wav_numpy.shape[0] // hop_length
-    if type_ == "dio":
-        f0, t = pyworld.dio(
-            wav_numpy.astype(np.double),
-            fs=sampling_rate,
-            f0_ceil=f0_max,
-            f0_floor=f0_min,
-            frame_period=1000 * hop_length / sampling_rate,
-        )
-    elif type_ == "harvest":
-        f0, t = pyworld.harvest(
-            wav_numpy.astype(np.double),
-            fs=sampling_rate,
-            f0_ceil=f0_max,
-            f0_floor=f0_min,
-            frame_period=1000 * hop_length / sampling_rate,
-        )
-    f0 = pyworld.stonemask(wav_numpy.astype(np.double), f0, t, sampling_rate)
-    for index, pitch in enumerate(f0):
-        f0[index] = round(pitch, 1)
-    return _resize_f0(f0, p_len)
-
-
-def compute_f0_crepe(
-    wav_numpy: ndarray[Any, dtype[float32]],
-    p_len: None | int = None,
-    sampling_rate: int = 44100,
-    hop_length: int = 512,
-    device: str = "cuda" if torch.cuda.is_available() else "cpu",
-    model: Literal["full", "tiny"] = "full",
-):
-    audio = torch.from_numpy(wav_numpy).to(device, copy=True)
-    audio = torch.unsqueeze(audio, dim=0)
-
-    if audio.ndim == 2 and audio.shape[0] > 1:
-        audio = torch.mean(audio, dim=0, keepdim=True).detach()
-    # (T) -> (1, T)
-    audio = audio.detach()
-
-    pitch: Tensor = torchcrepe.predict(
-        audio,
-        sampling_rate,
-        hop_length,
-        f0_min,
-        f0_max,
-        model,
-        batch_size=hop_length * 2,
-        device=device,
-        pad=True,
-    )
-
-    f0 = pitch.squeeze(0).cpu().numpy()
-    p_len = p_len or wav_numpy.shape[0] // hop_length
-    f0 = _resize_f0(f0, p_len)
-    return f0
-
-
-def compute_f0(
-    wav_numpy: ndarray[Any, dtype[float32]],
-    p_len: None | int = None,
-    sampling_rate: int = 44100,
-    hop_length: int = 512,
-    method: Literal["crepe", "crepe-tiny", "parselmouth", "dio", "harvest"] = "dio",
-    **kwargs,
-):
-    with timer() as t:
-        wav_numpy = wav_numpy.astype(np.float32)
-        wav_numpy /= np.quantile(np.abs(wav_numpy), 0.999)
-        if method in ["dio", "harvest"]:
-            f0 = compute_f0_pyworld(wav_numpy, p_len, sampling_rate, hop_length, method)
-        elif method == "crepe":
-            f0 = compute_f0_crepe(wav_numpy, p_len, sampling_rate, hop_length, **kwargs)
-        elif method == "crepe-tiny":
-            f0 = compute_f0_crepe(
-                wav_numpy, p_len, sampling_rate, hop_length, model="tiny", **kwargs
-            )
-        elif method == "parselmouth":
-            f0 = compute_f0_parselmouth(wav_numpy, p_len, sampling_rate, hop_length)
-        else:
-            raise ValueError(
-                "type must be dio, crepe, crepe-tiny, harvest or parselmouth"
-            )
-    rtf = t.elapsed / (len(wav_numpy) / sampling_rate)
-    LOG.info(f"F0 inference time:       {t.elapsed:.3f}s, RTF: {rtf:.3f}")
-    return f0
-
-
-def f0_to_coarse(f0: torch.Tensor | float):
-    is_torch = isinstance(f0, torch.Tensor)
-    f0_mel = 1127 * (1 + f0 / 700).log() if is_torch else 1127 * np.log(1 + f0 / 700)
-    f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * (f0_bin - 2) / (
-        f0_mel_max - f0_mel_min
-    ) + 1
-
-    f0_mel[f0_mel <= 1] = 1
-    f0_mel[f0_mel > f0_bin - 1] = f0_bin - 1
-    f0_coarse = (f0_mel + 0.5).long() if is_torch else np.rint(f0_mel).astype(np.int)
-    assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
-        f0_coarse.max(),
-        f0_coarse.min(),
-    )
-    return f0_coarse
-
-
 def download_file(
     url: str,
     filepath: Path | str,
-    chunk_size: int = 4 * 1024,
+    chunk_size: int = 64 * 1024,
     tqdm_cls: type = tqdm,
-    **kwargs,
+    skip_if_exists: bool = False,
+    overwrite: bool = False,
+    **tqdm_kwargs: Any,
 ):
+    if skip_if_exists is True and overwrite is True:
+        raise ValueError("skip_if_exists and overwrite cannot be both True")
     filepath = Path(filepath)
     filepath.parent.mkdir(parents=True, exist_ok=True)
     temppath = filepath.parent / f"{filepath.name}.download"
     if filepath.exists():
-        raise FileExistsError(f"{filepath} already exists")
+        if skip_if_exists:
+            return
+        elif not overwrite:
+            filepath.unlink()
+        else:
+            raise FileExistsError(f"{filepath} already exists")
     temppath.unlink(missing_ok=True)
     resp = requests.get(url, stream=True)
     total = int(resp.headers.get("content-length", 0))
-    with temppath.open("wb") as f, tqdm_cls(
-        total=total,
-        unit="iB",
-        unit_scale=True,
-        unit_divisor=1024,
-        **kwargs,
-    ) as pbar:
+    kwargs = (
+        dict(
+            total=total,
+            unit="iB",
+            unit_scale=True,
+            unit_divisor=1024,
+            desc=f"Downloading {filepath.name}",
+        )
+        | tqdm_kwargs
+    )
+    with temppath.open("wb") as f, tqdm_cls(**kwargs) as pbar:
         for data in resp.iter_content(chunk_size=chunk_size):
             size = f.write(data)
             pbar.update(size)
     temppath.rename(filepath)
 
 
-def ensure_pretrained_model(folder_path: Path, **kwargs) -> None:
-    model_urls = [
-        # "https://huggingface.co/innnky/sovits_pretrained/resolve/main/sovits4/G_0.pth",
-        "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/D_0.pth",
-        # "https://huggingface.co/innnky/sovits_pretrained/resolve/main/sovits4/D_0.pth",
-        "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/G_0.pth",
-    ]
-    for model_url in model_urls:
-        model_path = folder_path / model_url.split("/")[-1]
-        if not model_path.exists():
-            download_file(
-                model_url, model_path, desc=f"Downloading {model_path.name}", **kwargs
+PRETRAINED_MODEL_URLS = {
+    "hifi-gan": [
+        [
+            "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/D_0.pth",
+            "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/G_0.pth",
+        ],
+        [
+            "https://huggingface.co/Himawari00/so-vits-svc4.0-pretrain-models/resolve/main/D_0.pth",
+            "https://huggingface.co/Himawari00/so-vits-svc4.0-pretrain-models/resolve/main/G_0.pth",
+        ],
+    ],
+    "contentvec": [
+        [
+            "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/checkpoint_best_legacy_500.pt"
+        ],
+        [
+            "https://huggingface.co/Himawari00/so-vits-svc4.0-pretrain-models/resolve/main/checkpoint_best_legacy_500.pt"
+        ],
+        [
+            "http://obs.cstcloud.cn/share/obs/sankagenkeshi/checkpoint_best_legacy_500.pt"
+        ],
+    ],
+}
+from joblib import Parallel, delayed
+
+
+def ensure_pretrained_model(
+    folder_path: Path | str, type_: str, **tqdm_kwargs: Any
+) -> tuple[Path, ...] | None:
+    folder_path = Path(folder_path)
+    models_candidates = PRETRAINED_MODEL_URLS.get(type_, None)
+    if models_candidates is None:
+        LOG.warning(f"Unknown pretrained model type: {type_}")
+        return
+    for model_urls in models_candidates:
+        paths = [folder_path / model_url.split("/")[-1] for model_url in model_urls]
+        try:
+            Parallel(n_jobs=len(paths))(
+                [
+                    delayed(download_file)(
+                        url, path, position=i, skip_if_exists=True, **tqdm_kwargs
+                    )
+                    for i, (url, path) in enumerate(zip(model_urls, paths))
+                ]
             )
+            return tuple(paths)
+        except Exception as e:
+            LOG.exception(e)
+    return
 
 
-def ensure_hubert_model(**kwargs) -> Path:
-    vec_path = Path("checkpoint_best_legacy_500.pt")
-    vec_path.parent.mkdir(parents=True, exist_ok=True)
-    if not vec_path.exists():
-        # url = "http://obs.cstcloud.cn/share/obs/sankagenkeshi/checkpoint_best_legacy_500.pt"
-        # url = "https://huggingface.co/innnky/contentvec/resolve/main/checkpoint_best_legacy_500.pt"
-        url = "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/checkpoint_best_legacy_500.pt"
-        download_file(url, vec_path, desc="Downloading Hubert model", **kwargs)
-    return vec_path
-
-
-def get_hubert_model():
-    vec_path = ensure_hubert_model()
-    from fairseq import checkpoint_utils
+def get_hubert_model(device: torch.device) -> HubertModel:
+    (path,) = ensure_pretrained_model(Path("."), "contentvec")
 
     models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
-        [vec_path.as_posix()],
+        [path.as_posix()],
         suffix="",
     )
     model = models[0]
-    model.eval()
-    return model
-
-
-def get_hubert_content(hmodel, wav_16k_tensor):
-    with timer() as t:
-        feats = wav_16k_tensor
-        if feats.dim() == 2:  # double channels
-            feats = feats.mean(-1)
-        assert feats.dim() == 1, feats.dim()
-        feats = feats.view(1, -1)
-        padding_mask = torch.BoolTensor(feats.shape).fill_(False)
-        inputs = {
-            "source": feats.to(wav_16k_tensor.device),
-            "padding_mask": padding_mask.to(wav_16k_tensor.device),
-            "output_layer": 9,  # layer 9
-        }
-        with torch.no_grad():
-            logits = hmodel.extract_features(**inputs)
-            feats = hmodel.final_proj(logits[0])
-        res = feats.transpose(1, 2)
-    wav_len = wav_16k_tensor.shape[-1] / 16000
+    return model.eval().to(device)
+
+
+import warnings
+
+import torchaudio
+
+
+def get_content(
+    cmodel: HubertModel,
+    audio: torch.Tensor | ndarray[Any, Any],
+    device: torch.device | str,
+    sr: int,
+    legacy_final_proj: bool = False,
+) -> torch.Tensor:
+    audio = torch.as_tensor(audio)
+    if sr != HUBERT_SAMPLING_RATE:
+        audio = torchaudio.transforms.Resample(sr, HUBERT_SAMPLING_RATE)(audio).to(
+            device
+        )
+    if audio.ndim == 1:
+        audio = audio.unsqueeze(0)
+    with torch.no_grad(), timer() as t:
+        params = {"output_layer": 9} if legacy_final_proj else {}
+        c: torch.Tensor = cmodel.extract_features(audio, **params)[0]
+        if legacy_final_proj:
+            warnings.warn("legacy_final_proj is deprecated")
+            assert hasattr(cmodel, "final_proj")
+            assert isinstance(cmodel.final_proj, torch.nn.Module)
+            c = cmodel.final_proj(c)
+        c = c.transpose(1, 2)
+        # print(c.shape)
+    wav_len = audio.shape[-1] / HUBERT_SAMPLING_RATE
     LOG.info(
         f"HuBERT inference time  : {t.elapsed:.3f}s, RTF: {t.elapsed / wav_len:.3f}"
     )
-    return res
-
-
-def get_content(cmodel: Any, y: ndarray) -> ndarray:
-    with torch.no_grad():
-        c = cmodel.extract_features(y.squeeze(1))[0]
-    c = c.transpose(1, 2)
     return c
 
 
 def load_checkpoint(
-    checkpoint_path: Any,
-    model: Any,
-    optimizer: Any = None,
+    checkpoint_path: Path | str,
+    model: torch.nn.Module,
+    optimizer: torch.optim.Optimizer | None = None,
     skip_optimizer: bool = False,
-):
+) -> tuple[torch.nn.Module, torch.optim.Optimizer | None, float, int]:
     if not Path(checkpoint_path).is_file():
         raise FileNotFoundError(f"File {checkpoint_path} not found")
     checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
@@ -399,7 +181,11 @@ def load_checkpoint(
         and not skip_optimizer
         and checkpoint_dict["optimizer"] is not None
     ):
-        optimizer.load_state_dict(checkpoint_dict["optimizer"])
+        try:
+            optimizer.load_state_dict(checkpoint_dict["optimizer"])
+        except Exception as e:
+            LOG.exception(e)
+            LOG.warning("Failed to load optimizer state")
     saved_state_dict = checkpoint_dict["model"]
     if hasattr(model, "module"):
         state_dict = model.module.state_dict()
@@ -408,8 +194,6 @@ def load_checkpoint(
     new_state_dict = {}
     for k, v in state_dict.items():
         try:
-            # assert "dec" in k or "disc" in k
-            # print("load", k)
             new_state_dict[k] = saved_state_dict[k]
             assert saved_state_dict[k].shape == v.shape, (
                 saved_state_dict[k].shape,
@@ -428,7 +212,11 @@ def load_checkpoint(
 
 
 def save_checkpoint(
-    model, optimizer, learning_rate, iteration, checkpoint_path
+    model: torch.nn.Module,
+    optimizer: torch.optim.Optimizer,
+    learning_rate: float,
+    iteration: int,
+    checkpoint_path: Path | str,
 ) -> None:
     LOG.info(
         "Saving model and optimizer state at iteration {} to {}".format(
@@ -452,7 +240,7 @@ def save_checkpoint(
 
 def clean_checkpoints(
     path_to_models: Path | str, n_ckpts_to_keep: int = 2, sort_by_time: bool = True
-):
+) -> None:
     """Freeing up space by deleting saved ckpts
 
     Arguments:
@@ -492,15 +280,18 @@ def clean_checkpoints(
             to_delete.unlink()
 
 
+from torch.utils.tensorboard.writer import SummaryWriter
+
+
 def summarize(
-    writer,
-    global_step,
-    scalars={},
-    histograms={},
-    images={},
-    audios={},
-    audio_sampling_rate=22050,
-):
+    writer: SummaryWriter,
+    global_step: int,
+    scalars: dict[str, float] = {},
+    histograms: dict[str, ndarray] = {},
+    images: dict[str, ndarray] = {},
+    audios: dict[str, ndarray] = {},
+    audio_sampling_rate: int | None = None,
+) -> None:
     for k, v in scalars.items():
         writer.add_scalar(k, v, global_step)
     for k, v in histograms.items():
@@ -508,25 +299,22 @@ def summarize(
     for k, v in images.items():
         writer.add_image(k, v, global_step, dataformats="HWC")
     for k, v in audios.items():
+        if audio_sampling_rate is None:
+            raise ValueError("audio_sampling_rate must be provided")
         writer.add_audio(k, v, global_step, audio_sampling_rate)
 
 
-def latest_checkpoint_path(dir_path: Path | str, regex: str = "G_*.pth"):
+def latest_checkpoint_path(dir_path: Path | str, regex: str = "G_*.pth") -> Path | None:
     dir_path = Path(dir_path)
     name_key = lambda p: int(re.match(r"._(\d+)\.pth", p.name).group(1))
-    return list(sorted(dir_path.glob(regex), key=name_key))[-1]
-
-
-def plot_spectrogram_to_numpy(spectrogram):
-    global MATPLOTLIB_FLAG
-    if not MATPLOTLIB_FLAG:
-        import matplotlib
+    paths = list(sorted(dir_path.glob(regex), key=name_key))
+    if len(paths) == 0:
+        return None
+    return paths[-1]
 
-        matplotlib.use("Agg")
-        MATPLOTLIB_FLAG = True
-    import matplotlib.pylab as plt
-    import numpy as np
 
+def plot_spectrogram_to_numpy(spectrogram: ndarray) -> ndarray:
+    matplotlib.use("Agg")
     fig, ax = plt.subplots(figsize=(10, 2))
     im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
     plt.colorbar(im, ax=ax)
@@ -541,7 +329,7 @@ def plot_spectrogram_to_numpy(spectrogram):
     return data
 
 
-def load_wav_to_torch(full_path: Path | str):
+def load_wav_to_torch(full_path: Path | str) -> tuple[torch.Tensor, int]:
     sampling_rate, data = read(full_path)
     return torch.FloatTensor(data.astype(np.float32)), sampling_rate
 
@@ -552,7 +340,9 @@ def load_filepaths_and_text(filename: Path | str, split="|"):
     return filepaths_and_text
 
 
-def get_hparams(config_path: Path, model_path: Path, init: bool = True) -> HParams:
+def get_backup_hparams(
+    config_path: Path, model_path: Path, init: bool = True
+) -> HParams:
     model_path.mkdir(parents=True, exist_ok=True)
     config_save_path = model_path / "config.json"
     if init:
@@ -570,18 +360,17 @@ def get_hparams(config_path: Path, model_path: Path, init: bool = True) -> HPara
     return hparams
 
 
-def get_hparams_from_file(config_path: Path | str) -> HParams:
+def get_hparams(config_path: Path | str) -> HParams:
     config = json.loads(Path(config_path).read_text())
     hparams = HParams(**config)
     return hparams
 
 
-def repeat_expand_2d(content: ndarray, target_len: int) -> ndarray:
+def repeat_expand_2d(content: torch.Tensor, target_len: int) -> torch.Tensor:
     # content : [h, t]
-
     src_len = content.shape[-1]
-    target = torch.zeros([content.shape[0], target_len], dtype=torch.float).to(
-        content.device
+    target = torch.zeros(
+        [content.shape[0], target_len], dtype=content.dtype, device=content.device
     )
     temp = torch.arange(src_len + 1) * target_len / src_len
     current_pos = 0
@@ -591,37 +380,31 @@ def repeat_expand_2d(content: ndarray, target_len: int) -> ndarray:
         else:
             current_pos += 1
             target[:, i] = content[:, current_pos]
-
     return target
 
 
-class HParams:
-    def __init__(self, **kwargs):
-        for k, v in kwargs.items():
-            if type(v) == dict:
-                v = HParams(**v)
-            self[k] = v
-
-    def keys(self):
-        return self.__dict__.keys()
-
-    def items(self):
-        return self.__dict__.items()
-
-    def values(self):
-        return self.__dict__.values()
+def plot_data_to_numpy(x: ndarray, y: ndarray) -> ndarray:
+    matplotlib.use("Agg")
+    fig, ax = plt.subplots(figsize=(10, 2))
+    plt.plot(x)
+    plt.plot(y)
+    plt.tight_layout()
 
-    def __len__(self):
-        return len(self.__dict__)
+    fig.canvas.draw()
+    data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+    plt.close()
+    return data
 
-    def __getitem__(self, key):
-        return getattr(self, key)
 
-    def __setitem__(self, key, value):
-        return setattr(self, key, value)
+def get_gpu_memory(type_: Literal["total", "free", "used"]) -> Sequence[int]:
+    command = f"nvidia-smi --query-gpu=memory.{type_} --format=csv"
+    memory_free_info = (
+        subprocess.check_output(command.split()).decode("ascii").split("\n")[:-1][1:]
+    )
+    memory_free_values = [int(x.split()[0]) for i, x in enumerate(memory_free_info)]
+    return memory_free_values
 
-    def __contains__(self, key):
-        return key in self.__dict__
 
-    def __repr__(self):
-        return self.__dict__.__repr__()
+def get_total_gpu_memory(type_: Literal["total", "free", "used"]) -> int:
+    return sum(get_gpu_memory(type_))
diff --git a/src/so_vits_svc_fork/vdecoder/__init__.py b/src/so_vits_svc_fork/vdecoder/__init__.py
deleted file mode 100644
index e69de29b..00000000
diff --git a/src/so_vits_svc_fork/vdecoder/hifigan/__init__.py b/src/so_vits_svc_fork/vdecoder/hifigan/__init__.py
deleted file mode 100644
index e69de29b..00000000
diff --git a/src/so_vits_svc_fork/vdecoder/hifigan/env.py b/src/so_vits_svc_fork/vdecoder/hifigan/env.py
deleted file mode 100644
index a1231a2d..00000000
--- a/src/so_vits_svc_fork/vdecoder/hifigan/env.py
+++ /dev/null
@@ -1,15 +0,0 @@
-import os
-import shutil
-
-
-class AttrDict(dict):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.__dict__ = self
-
-
-def build_env(config, config_name, path):
-    t_path = os.path.join(path, config_name)
-    if config != t_path:
-        os.makedirs(path, exist_ok=True)
-        shutil.copyfile(config, os.path.join(path, config_name))
diff --git a/src/so_vits_svc_fork/vdecoder/hifigan/nvSTFT.py b/src/so_vits_svc_fork/vdecoder/hifigan/nvSTFT.py
deleted file mode 100644
index a33805d6..00000000
--- a/src/so_vits_svc_fork/vdecoder/hifigan/nvSTFT.py
+++ /dev/null
@@ -1,162 +0,0 @@
-import os
-
-os.environ["LRU_CACHE_CAPACITY"] = "3"
-from logging import getLogger
-
-import librosa
-import numpy as np
-import soundfile as sf
-import torch
-import torch.utils.data
-from librosa.filters import mel as librosa_mel_fn
-
-LOG = getLogger(__name__)
-
-
-def load_wav_to_torch(full_path, target_sr=None, return_empty_on_exception=False):
-    sampling_rate = None
-    try:
-        data, sampling_rate = sf.read(full_path, always_2d=True)  # than soundfile.
-    except Exception as ex:
-        LOG.info(f"'{full_path}' failed to load.\nException:")
-        LOG.info(ex)
-        if return_empty_on_exception:
-            return [], sampling_rate or target_sr or 32000
-        else:
-            raise Exception(ex)
-
-    if len(data.shape) > 1:
-        data = data[:, 0]
-        assert (
-            len(data) > 2
-        )  # check duration of audio file is > 2 samples (because otherwise the slice operation was on the wrong dimension)
-
-    if np.issubdtype(data.dtype, np.integer):  # if audio data is type int
-        max_mag = -np.iinfo(
-            data.dtype
-        ).min  # maximum magnitude = min possible value of intXX
-    else:  # if audio data is type fp32
-        max_mag = max(np.amax(data), -np.amin(data))
-        max_mag = (
-            (2**31) + 1
-            if max_mag > (2**15)
-            else ((2**15) + 1 if max_mag > 1.01 else 1.0)
-        )  # data should be either 16-bit INT, 32-bit INT or [-1 to 1] float32
-
-    data = torch.FloatTensor(data.astype(np.float32)) / max_mag
-
-    if (
-        torch.isinf(data) | torch.isnan(data)
-    ).any() and return_empty_on_exception:  # resample will crash with inf/NaN inputs. return_empty_on_exception will return empty arr instead of except
-        return [], sampling_rate or target_sr or 32000
-    if target_sr is not None and sampling_rate != target_sr:
-        data = torch.from_numpy(
-            librosa.core.resample(
-                data.numpy(), orig_sr=sampling_rate, target_sr=target_sr
-            )
-        )
-        sampling_rate = target_sr
-
-    return data, sampling_rate
-
-
-def dynamic_range_compression(x, C=1, clip_val=1e-5):
-    return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
-
-
-def dynamic_range_decompression(x, C=1):
-    return np.exp(x) / C
-
-
-def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
-    return torch.log(torch.clamp(x, min=clip_val) * C)
-
-
-def dynamic_range_decompression_torch(x, C=1):
-    return torch.exp(x) / C
-
-
-class STFT:
-    def __init__(
-        self,
-        sr=22050,
-        n_mels=80,
-        n_fft=1024,
-        win_size=1024,
-        hop_length=256,
-        fmin=20,
-        fmax=11025,
-        clip_val=1e-5,
-    ):
-        self.target_sr = sr
-
-        self.n_mels = n_mels
-        self.n_fft = n_fft
-        self.win_size = win_size
-        self.hop_length = hop_length
-        self.fmin = fmin
-        self.fmax = fmax
-        self.clip_val = clip_val
-        self.mel_basis = {}
-        self.hann_window = {}
-
-    def get_mel(self, y, center=False):
-        sampling_rate = self.target_sr
-        n_mels = self.n_mels
-        n_fft = self.n_fft
-        win_size = self.win_size
-        hop_length = self.hop_length
-        fmin = self.fmin
-        fmax = self.fmax
-        clip_val = self.clip_val
-
-        if torch.min(y) < -1.0:
-            LOG.info("min value is ", torch.min(y))
-        if torch.max(y) > 1.0:
-            LOG.info("max value is ", torch.max(y))
-
-        if fmax not in self.mel_basis:
-            mel = librosa_mel_fn(
-                sr=sampling_rate, n_fft=n_fft, n_mels=n_mels, fmin=fmin, fmax=fmax
-            )
-            self.mel_basis[str(fmax) + "_" + str(y.device)] = (
-                torch.from_numpy(mel).float().to(y.device)
-            )
-            self.hann_window[str(y.device)] = torch.hann_window(self.win_size).to(
-                y.device
-            )
-
-        y = torch.nn.functional.pad(
-            y.unsqueeze(1),
-            (int((n_fft - hop_length) / 2), int((n_fft - hop_length) / 2)),
-            mode="reflect",
-        )
-        y = y.squeeze(1)
-
-        spec = torch.stft(
-            y,
-            n_fft,
-            hop_length=hop_length,
-            win_length=win_size,
-            window=self.hann_window[str(y.device)],
-            center=center,
-            pad_mode="reflect",
-            normalized=False,
-            onesided=True,
-        )
-        # LOG.info(111,spec)
-        spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
-        # LOG.info(222,spec)
-        spec = torch.matmul(self.mel_basis[str(fmax) + "_" + str(y.device)], spec)
-        # LOG.info(333,spec)
-        spec = dynamic_range_compression_torch(spec, clip_val=clip_val)
-        # LOG.info(444,spec)
-        return spec
-
-    def __call__(self, audiopath):
-        audio, sr = load_wav_to_torch(audiopath, target_sr=self.target_sr)
-        spect = self.get_mel(audio.unsqueeze(0)).squeeze(0)
-        return spect
-
-
-stft = STFT()
diff --git a/src/so_vits_svc_fork/vdecoder/hifigan/utils.py b/src/so_vits_svc_fork/vdecoder/hifigan/utils.py
deleted file mode 100644
index 5839f9e4..00000000
--- a/src/so_vits_svc_fork/vdecoder/hifigan/utils.py
+++ /dev/null
@@ -1,71 +0,0 @@
-import glob
-import os
-from logging import getLogger
-
-# matplotlib.use("Agg")
-import matplotlib.pylab as plt
-import torch
-from torch.nn.utils import weight_norm
-
-LOG = getLogger(__name__)
-
-
-def plot_spectrogram(spectrogram):
-    fig, ax = plt.subplots(figsize=(10, 2))
-    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
-    plt.colorbar(im, ax=ax)
-
-    fig.canvas.draw()
-    plt.close()
-
-    return fig
-
-
-def init_weights(m, mean=0.0, std=0.01):
-    classname = m.__class__.__name__
-    if classname.find("Conv") != -1:
-        m.weight.data.normal_(mean, std)
-
-
-def apply_weight_norm(m):
-    classname = m.__class__.__name__
-    if classname.find("Conv") != -1:
-        weight_norm(m)
-
-
-def get_padding(kernel_size, dilation=1):
-    return int((kernel_size * dilation - dilation) / 2)
-
-
-def load_checkpoint(filepath, device):
-    assert os.path.isfile(filepath)
-    LOG.info(f"Loading '{filepath}'")
-    checkpoint_dict = torch.load(filepath, map_location=device)
-    LOG.info("Complete.")
-    return checkpoint_dict
-
-
-def save_checkpoint(filepath, obj):
-    LOG.info(f"Saving checkpoint to {filepath}")
-    torch.save(obj, filepath)
-    LOG.info("Complete.")
-
-
-def del_old_checkpoints(cp_dir, prefix, n_models=2):
-    pattern = os.path.join(cp_dir, prefix + "????????")
-    cp_list = glob.glob(pattern)  # get checkpoint paths
-    cp_list = sorted(cp_list)  # sort by iter
-    if len(cp_list) > n_models:  # if more than n_models models are found
-        for cp in cp_list[
-            :-n_models
-        ]:  # delete the oldest models other than last n_models
-            open(cp, "w").close()  # empty file contents
-            os.unlink(cp)  # delete file (move to trash when using Colab)
-
-
-def scan_checkpoint(cp_dir, prefix):
-    pattern = os.path.join(cp_dir, prefix + "????????")
-    cp_list = glob.glob(pattern)
-    if len(cp_list) == 0:
-        return None
-    return sorted(cp_list)[-1]
diff --git a/tests/test_main.py b/tests/test_main.py
index dadc4b3f..36985dbf 100644
--- a/tests/test_main.py
+++ b/tests/test_main.py
@@ -10,41 +10,49 @@
 class TestMain(TestCase):
     def test_import(self):
         import so_vits_svc_fork.cluster.train_cluster  # noqa
-        import so_vits_svc_fork.inference_main  # noqa
-        import so_vits_svc_fork.onnx_export  # noqa
-        import so_vits_svc_fork.preprocess_flist_config  # noqa
-        import so_vits_svc_fork.preprocess_hubert_f0  # noqa
-        import so_vits_svc_fork.preprocess_resample  # noqa
-        import so_vits_svc_fork.preprocess_split  # noqa
+        import so_vits_svc_fork.inference.main  # noqa
+
+        # import so_vits_svc_fork.modules.onnx._export  # noqa
+        import so_vits_svc_fork.preprocessing.preprocess_flist_config  # noqa
+        import so_vits_svc_fork.preprocessing.preprocess_hubert_f0  # noqa
+        import so_vits_svc_fork.preprocessing.preprocess_resample  # noqa
+        import so_vits_svc_fork.preprocessing.preprocess_split  # noqa
         import so_vits_svc_fork.train  # noqa
 
     def test_infer(self):
         if IS_CI:
             raise SkipTest("Skip inference test on CI")
-        from so_vits_svc_fork.inference_main import infer  # noqa
+        from so_vits_svc_fork.inference.main import infer  # noqa
 
         # infer("tests/dataset_raw/34j/1.wav", "tests/configs/config.json", "tests/logs/44k")
 
     def test_preprocess(self):
-        from so_vits_svc_fork.preprocess_resample import preprocess_resample
+        from so_vits_svc_fork.preprocessing.preprocess_resample import (
+            preprocess_resample,
+        )
 
         preprocess_resample(
             "tests/dataset_raw", "tests/dataset/44k", 44100, n_jobs=1 if IS_CI else -1
         )
 
-        from so_vits_svc_fork.preprocess_flist_config import preprocess_config
+        from so_vits_svc_fork.preprocessing.preprocess_flist_config import (
+            preprocess_config,
+        )
 
         preprocess_config(
             "tests/dataset/44k",
             "tests/filelists/train.txt",
             "tests/filelists/val.txt",
             "tests/filelists/test.txt",
-            "tests/configs/config.json",
+            "tests/configs/44k/config.json",
+            "so-vits-svc-4.0v1",
         )
 
         if IS_CI:
             raise SkipTest("Skip hubert and f0 test on CI")
-        from so_vits_svc_fork.preprocess_hubert_f0 import preprocess_hubert_f0
+        from so_vits_svc_fork.preprocessing.preprocess_hubert_f0 import (
+            preprocess_hubert_f0,
+        )
 
         preprocess_hubert_f0("tests/dataset/44k", "tests/configs/44k/config.json")