GTCRN-Light v3

Lightweight, deployment-friendly reimplementation of GTCRN for single-channel speech enhancement.

（中文简介）GTCRN-Light v3（下文简称 v3）是在官方 GTCRN 代码基础上做的等价轻量化版本：只改算子、不改语义，保持输入输出接口完全兼容，方便直接替换到现有 SEtrain / TRT-SE 等流水线中。

1. Introduction

GTCRN-Light v3 is a structurally equivalent, heavily compressed version of the original Grouped Temporal Convolutional Recurrent Network (GTCRN) for speech enhancement.
Instead of changing the task setting or loss, v3 keeps:

• complex ratio mask (CRM) prediction in the ERB domain,
• the encoder → DPGRNN → decoder pipeline,
• the overall STFT → ERB → enhancement → iERB → iSTFT data flow,

and only replaces heavy building blocks with lighter operators.

Design goals:

• Keep the same input / output interface as the official GTCRN implementation.
• Reduce parameters and MACs while preserving DPGRNN’s long-term time–frequency modelling.
• Maintain strictly aligned frequency shapes (e.g. 129 → 65 → 33) for stable deployment on edge devices.

1.1 Complexity & Example Metrics

Complexity for a 1 s, 16 kHz input (n_fft=512, hop=256, win=512):

Model	Params (M)	MACs (G/s)	Note
GTCRN (official)	≈0.048	≈0.033	From the original repo
GTCRN-Light v3	0.008	0.017	Spec core, this repository

Example evaluation (one internal experiment, VCTK-DEMAND style setting):

• PESQ (wb): 2.5966
• STOI: 0.9338
• ESTOI: 0.8299
• SI-SNR: 17.89 dB

Numbers above are for reference only; your results will depend on the dataset and training recipe.

2. Features

• Operator-level lightweighting
  Depthwise separable convolutions, bottlenecked RNNs and lightweight temporal gating drastically reduce parameters and MACs.

• Semantic equivalence to GTCRN
  Preserves ERB subband processing, CRM prediction and dual-path GRNN ordering; no shortcut in modelling capability.

• Shape-stable encoder / decoder
  Integer frequency down/up sampling (129 → 65 → 33 and back) avoids odd–even drift and keeps skip connections aligned.

• Deployment friendly
  Fewer stateful components, no exotic ops, and an ERB projection implemented as a fixed buffer, which makes exporting / quantising easier.

For more design details and training notes (Chinese), please refer to 轻量化文档.md.

3. Environment & Installation

Tested with:

• Python 3.10
• PyTorch 2.0.1 + CUDA 11.8
• Linux x86_64

Minimal dependencies:

conda create -n gtcrn_light python=3.10
conda activate gtcrn_light

pip install torch torchaudio librosa numpy scipy soundfile tqdm pyyaml
# Optional: for evaluation
pip install pesq pystoi

You can of course adapt the environment to your own infrastructure.

4. Repository Layout

A minimal layout is expected as:

.
├── configs/
│   └── cfg_train_gtcrn_spec.yaml
├── models/
│   └── gtrcn_light_v3.py
├── tools/
│   ├── batch_infer_gtcrn_spec_light.py
│   ├── make_pairs_from_outputs.py
│   └── print_model_stats.py
├── train_my_spec_light_v3.py
├── 轻量化文档.md
└── README.md

Your actual project may contain more scripts (data preparation, logging, etc.); adjust accordingly.

5. Training

The main training script for the spectrogram-domain model is:

python train_my_spec_light_v3.py -C configs/cfg_train_gtcrn_spec.yaml -D 0

Typical configuration items in cfg_train_gtcrn_spec.yaml include:

• Dataset paths for noisy / clean speech
• STFT parameters: sr=16000, n_fft=512, hop=256, win=512
• Optimiser and LR schedule
• GTCRN-Light specific knobs: width_mult, bottleneck ratio r, use_two_dpgrnn, rnn_bidirectional

See the comments inside your config file for details.

5.1 Training Tips (from practice)

A few practical suggestions when training v3:

• Use combined losses on real/imaginary parts, magnitude, and optionally phase or waveform-based criteria (e.g. SI-SDR).
• Warm up the learning rate for a few epochs and apply gradient clipping (e.g. 5.0) to stabilise DPGRNN.
• SpecAug-style light masking in time / frequency often improves robustness.
• If you already have a strong GTCRN model, you can distil intermediate features from it to accelerate convergence.

More detailed recipes are summarised in 轻量化文档.md.

6. Inference

6.1 Batch inference (spectrogram version)

python tools/batch_infer_gtcrn_spec_light.py   --ckpt /path/to/your/best_model_xxx.tar   --in_dir  ../data/test/noisy   --out_dir ../SEtrain/outs/gtcrn_spec_test_light3   --sr 16000 --n_fft 512 --hop 256 --win 512 --device cuda

This script reads all noisy .wav files from --in_dir, applies the model and writes enhanced files to --out_dir.

6.2 Build paired list for evaluation

python tools/make_pairs_from_outputs.py   --enh_dir  ../SEtrain/outs/gtcrn_spec_test_light3   --clean_dir ../data/test/clean   --out pairs_gtcrn_light_v3.txt

The generated text file lists (enhanced, clean) pairs and can be used by your metric script.

6.3 Model complexity

Before running the complexity script, temporarily ensure that the class name in models/gtrcn_light_v3.py matches what print_model_stats.py expects (for example, rename GTCRN_light_v3 to GTCRN_light).

python tools/print_model_stats.py   --mode spec   --arch gtrcn_light   --model-path models/gtrcn_light_v3.py   --width-mult 1.0   --two-rnn   --rnn-bidir

The script reports parameter count and MACs for a 1-second input:

=== GTCRN-Light v3 (Spec Core) ===
Config  : width_mult=1.0, use_two_dpgrnn=True, rnn_bidirectional=True
Input(1s): F=257, T=63  (n_fft=512, hop=256, win=512)
Para. (M):  0.008
MACs (G/s): 0.017  # core network, without STFT/ISTFT

7. Design Highlights

Very briefly, GTCRN-Light v3 introduces:

• GT-ConvLite
  Depthwise separable temporal convolutions followed by pointwise projection and lightweight gating, providing a GT-Conv-like receptive field with far fewer parameters.

• TRALite
  RNN-free temporal energy gating implemented with 1D depthwise + pointwise convolutions on frame-wise energy trajectories.

• Bottlenecked DPGRNN
  Intra / inter RNN paths operate in a low-rank space C → r → C, greatly reducing parameters while preserving dual-path modelling.

• Fixed ERB projection
  ERB mapping is implemented as a fixed buffer instead of learnable Linear layers, cutting parameters and easing deployment.

See 轻量化文档.md for diagrams and a more detailed discussion.

8. Acknowledgements

This repository is a lightweight, engineering-oriented reimplementation built on top of the official GTCRN project:

• Xiaobin Rong et al., “GTCRN: A Speech Enhancement Model Requiring Ultralow Computational Resources,” ICASSP 2024.
• Official code repository: https://github.com/Xiaobin-Rong/gtcrn

9. Citation

This work main thanks to https://github.com/lililili60

If you use this repo in academic work, please consider citing both the original GTCRN paper and this repository.