Add greedy search and modified_beam_search for non-streaming ASR (#95)

* WIP: Add greedy search for non-streaming ASR

* Use kaldi_native_io to read wave samples

* First working version of greedy search for offline ASR.

* Refactor the code

* Support decoding from wav.scp and feats.scp

* Update CI

* Add doc for offline ASR with sherpa's C++ frontend.

* remove sentencepiece

* Fix style issues

* support conda install

* build for windows

* Add FAQs for installation

* Add version info

* Fix building on macOS

* Add more doc

* Typo fixes

.flake8

@@ -11,6 +11,7 @@ per-file-ignores =
 exclude =
   .git,
   ./cmake,
+  ./scripts,
   ./triton,
   ./sherpa/python/sherpa/__init__.py,
   ./sherpa/python/sherpa/decode.py,

.github/scripts/generate_feats_scp.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+"""
+Usage:
+
+    ./generate_feats_scp.py scp:wav.scp ark,scp:feats.ark,feats.scp
+
+It generates `feats.ark` and `feats.scp` from `wav.scp`.
+
+Different from Kaldi's `compute-fbank-feats`, this scripts uses
+normalized samples in the range [-1, 1] to compute features.
+"""
+
+import sys
+
+import kaldi_native_io as kio
+import kaldifeat
+import torch
+
+
+def main():
+    rspecifier = sys.argv[1]
+    wspecifier = sys.argv[2]
+
+    opts = kaldifeat.FbankOptions()
+    opts.frame_opts.dither = 0
+    opts.mel_opts.num_bins = 80
+    fbank = kaldifeat.Fbank(opts)
+
+    with kio.SequentialWaveReader(rspecifier) as ki:
+        with kio.FloatMatrixWriter(wspecifier) as ko:
+            for key, value in ki:
+                tensor = torch.from_numpy(value.data.numpy()).clone().squeeze(0)
+                tensor = tensor / 32768
+                features = fbank(tensor)
+                ko.write(key, features.numpy())
+
+
+if __name__ == "__main__":
+    main()

.github/scripts/generate_wav_scp.sh

@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+cat > wav.scp <<EOF
+wav1 icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/test_wavs/1089-134686-0001.wav
+wav2 icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/test_wavs/1221-135766-0001.wav
+wav3 icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/test_wavs/1221-135766-0002.wav
+EOF

.github/scripts/generate_wav_scp_aishell.sh

@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+cat > wav_aishell.scp <<EOF
+wav1 icefall-aishell-pruned-transducer-stateless3-2022-06-20/test_wavs/BAC009S0764W0121.wav
+wav2 icefall-aishell-pruned-transducer-stateless3-2022-06-20/test_wavs/BAC009S0764W0122.wav
+wav3 icefall-aishell-pruned-transducer-stateless3-2022-06-20/test_wavs/BAC009S0764W0123.wav
+EOF

.github/workflows/build-conda-cpu.yaml

@@ -0,0 +1,133 @@
+name: build_conda_cpu
+
+on:
+  push:
+    tags:
+      - '*'
+
+jobs:
+  generate_build_matrix:
+    # see https://github.com/pytorch/pytorch/pull/50633
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+      - name: Generating build matrix
+        id: set-matrix
+        run: |
+          # outputting for debugging purposes
+          python scripts/github_actions/generate_build_matrix.py
+          MATRIX=$(python scripts/github_actions/generate_build_matrix.py)
+          echo "::set-output name=matrix::${MATRIX}"
+
+  build_conda_cpu:
+    needs: generate_build_matrix
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        ${{ fromJson(needs.generate_build_matrix.outputs.matrix) }}
+
+    steps:
+      # refer to https://github.com/actions/checkout
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - uses: conda-incubator/setup-miniconda@v2
+        with:
+          auto-update-conda: true
+          python-version: ${{ matrix.python-version }}
+          activate-environment: sherpa
+
+      - name: Install conda dependencies
+        shell: bash -l {0}
+        run: |
+          conda install -y -q anaconda-client
+          conda install -y -q conda-build
+          conda install -y -q -c k2-fsa -c kaldifeat -c kaldi_native_io -c pytorch k2 kaldifeat kaldi_native_io pytorch=${{ matrix.torch }} cpuonly
+
+      - name: Display MKL
+        if: startsWith(matrix.os, 'macos') || startsWith(matrix.os, 'ubuntu')
+        shell: bash -l {0}
+        run: |
+          ls -lh $CONDA_PREFIX/lib/libmkl*
+
+      - name: Display Python version
+        shell: bash -l {0}
+        run: |
+          python -c "import sys; print(sys.version)"
+          which python
+
+      - name: Display conda info
+        shell: bash -l {0}
+        run: |
+          conda env list
+          conda info
+          which conda
+          python --version
+          which python
+          python -m torch.utils.collect_env
+
+      - name: Build sherpa
+        if: startsWith(matrix.os, 'ubuntu') || startsWith(matrix.os, 'macos')
+        shell: bash -l {0}
+        env:
+          SHERPA_PYTHON_VERSION: ${{ matrix.python-version}}
+          SHERPA_TORCH_VERSION: ${{ matrix.torch }}
+          SHERPA_CONDA_TOKEN: ${{ secrets.SHERPA_CONDA_TOKEN}}
+        run: |
+          ./scripts/build_conda_cpu.sh
+
+      - name: Build sherpa
+        if: startsWith(matrix.os, 'windows')
+        shell: bash -l {0}
+        env:
+          SHERPA_PYTHON_VERSION: ${{ matrix.python-version}}
+          SHERPA_TORCH_VERSION: ${{ matrix.torch }}
+          SHERPA_CONDA_TOKEN: ${{ secrets.SHERPA_CONDA_TOKEN}}
+        run: |
+          # ./scripts/build_conda_cpu_windows.sh
+          ./scripts/build_conda_cpu.sh
+
+      - name: Display generated files
+        if: startsWith(matrix.os, 'ubuntu')
+        run: |
+          ls -lh /usr/share/miniconda/envs/sherpa/conda-bld/linux-64
+
+      - name: Upload generated files
+        if: startsWith(matrix.os, 'ubuntu')
+        uses: actions/upload-artifact@v2
+        with:
+          name: cpu-torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-${{ matrix.os }}
+          path: /usr/share/miniconda/envs/sherpa/conda-bld/linux-64/*.tar.bz2
+
+      - name: Display generated files
+        if: startsWith(matrix.os, 'windows')
+        shell: bash -l {0}
+        run: |
+          ls -lh /c/Miniconda/envs/sherpa/conda-bld
+          ls -lh /c/Miniconda/envs/sherpa/conda-bld/*/*
+          ls -lh /c/Miniconda/envs/sherpa/conda-bld/win-64/*
+
+      - name: Upload generated files
+        if: startsWith(matrix.os, 'windows')
+        uses: actions/upload-artifact@v2
+        with:
+          name: cpu-torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-${{ matrix.os }}
+          path: c:/Miniconda/envs/sherpa/conda-bld/win-64/*.tar.bz2
+
+      - name: Display generated files
+        if: startsWith(matrix.os, 'macos')
+        run: |
+          ls -lh /usr/local/miniconda/envs/sherpa/conda-bld/osx-64
+
+      - name: Upload generated files
+        if: startsWith(matrix.os, 'macos')
+        uses: actions/upload-artifact@v2
+        with:
+          name: cpu-torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-${{ matrix.os }}
+          path: /usr/local/miniconda/envs/sherpa/conda-bld/osx-64/*.tar.bz2

.github/workflows/build-doc.yml

@@ -56,7 +56,7 @@ jobs:
       - name: Install PyTorch ${{ matrix.torch }}
         shell: bash
         run: |
-          python3 -m pip install -qq --upgrade pip
+          python3 -m pip install -qq --upgrade pip kaldi_native_io
           python3 -m pip install -qq wheel twine typing_extensions websockets sentencepiece>=0.1.96 soundfile
           python3 -m pip install -qq torch==${{ matrix.torch }}+cpu numpy -f https://download.pytorch.org/whl/torch_stable.html