TAAnet

TAANet

A PyTorch implementation of TAANet on WSJ0-2mix described in the paper "Time-domain Adaptive Attention Network for Single-channel Speech Separation".

This implementation is based on:

1. https://github.com/kaituoxu/Conv-TasNet,

2. https://github.com/yluo42/TAC.

Thanks Kaituo and Yi Luo for sharing.

---

step 1：

Generate .json files with wav path and length.

python ./utils/preprocess.py

step 2：

Train the model with train & valid set.

CUDA_VISIBLE_DEVICES=0 python train_taanet.py

step 3：

Evaluate the trained model with test set.

CUDA_VISIBLE_DEVICES=0 python evaluate_taanet.py

We obtain a SI-SNRi of 20.7 dB and SDRi of 20.9 dB on WSJ0-2mix test set, and the model is available at saves/temp/temp_best.pth.tar.

step 4：

Separate the mixed speech.

CUDA_VISIBLE_DEVICES=0 python separate_taanet.py

Some separation samples can be found here.

---

Please cite the following reference if you utilize this repository for your project.

@article{cai2022,
author = {Cai, Jingxiang and Wang, Kunpeng and Yao, Juan and Zhou, Hao},
title = {Time-domain Adaptive Attention Network for Single-channel Speech Separation},
journal = {},
year = {2022}，
pages = {}
}