TENET: A Time-reversal Enhancement Network for noise-robust ASR (ASRU 2021)

A PyTorch implementation of our paper "TENET: A Time-reversal Enhancement Network for noise-robust ASR".

• Audio samples can be found from here.

• Pretrained model weights can be downloaded from here.

Requirements

• See requirements.txt

Usage

Inference

Usage: ./inference.sh <noisy.scp> <cpt-dir> <dump-dir>
  option: --ref-scp(given clean counterparts and calculate pesq/sisnr metrics) 
          --remove-wav(false)
          --model(TENET)
          --gpu (0) 
          --fs (16000) 
          --nj (1)

Training from scratch

1. Download pretrained wav2vec model and put it in pretrain/wav2vec_large.pt

2. Configure training settings and model hyperparameters from nnet/conf.py.

3. Full experiment command:

Usage: ./train.sh <cpt-dir> <model> <exp-id>
  egs: ./train.sh exp/voicebank TENET(TCN-skip, DCCRN, PFPL, DPTNet) freqdpt_base
  options: --resume (path/to/best.pt.tar) 
           --gpu (0)
           --epochs (100)
           --batch-size (4) 
           --cache-size (10)

Reference

This repository contains codes from:

• ConvTasNet - https://github.com/funcwj/conv-tasnet
• DPTNet - https://github.com/asteroid-team/asteroid
• DCCRN - https://github.com/huyanxin/DeepComplexCRN
• PFPL - https://github.com/aleXiehta/PhoneFortifiedPerceptualLoss
• SETK toolkit - https://github.com/funcwj/setk

Citation

If you find this repository useful, please cite the following paper:

@inproceedings{chao2021tenet,
  title = {TENET: A Time-reversal Enhancement Network for noise-robust ASR},
  author = {Fu-An Chao and Shao-Wei Fan Jiang and Bi-Cheng Yan 
            and Jeih-weih Hung and Berlin Chen},
  booktitle = {2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  year = {2021},
}