Dual-path-RNN-Pytorch

Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation implemented by Pytorch

If you have any questions, you can ask them through the issue.

If you find this project helpful, you can give me a star generously.

Demo Pages: Results of pure speech separation model

Plan

• 2020-02-01: Reading article “Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation”. Zhihu Article link "阅读笔记”Dual-path RNN for Speech Separation“". Blog Article link "阅读笔记《Dual-path RNN for speech separation》". Both articles are interpretations of the paper. If you have any questions, welcome to discuss with me

• 2020-02-02: Complete data preprocessing, data set code. Dataset Code: /data_loader/Dataset.py

• 2020-02-03: Complete Conv-TasNet Framework (Update /model/model.py, Trainer_Tasnet.py, Train_Tasnet.py)

• 2020-02-07: Complete Training code. (Update /model/model_rnn.py) and Test parameters and some details are being adjusted.

• 2020-02-08: Fixed the code's bug.

• 2020-02-11: Complete Testing code.

Dataset

We used the WSJ0 dataset as our training, test, and validation sets. Below is the data download link and mixed audio code for WSJ0.

• Audio mix Sample

• WSJ0 Dataset

• Create Dataset

Training

Training for Conv-TasNet model

1. First, you need to generate the scp file using the following command. The content of the scp file is "filename && path".

python create_scp.py

2. Then you can modify the training and model parameters through "config/Conv_Tasnet/train.yml".

cd config/Conv-Tasnet
vim train.yml

3. Then use the following command in the root directory to train the model.

python train_Tasnet.py --opt config/Conv_Tasnet/train.yml

Training for Dual Path RNN model

1. First, you need to generate the scp file using the following command. The content of the scp file is "filename && path".

python create_scp.py

2. Then you can modify the training and model parameters through "config/Dual_RNN/train.yml".

cd config/Dual_RNN
vim train.yml

3. Then use the following command in the root directory to train the model.

python train_rnn.py --opt config/Dual_RNN/train.yml

Inference

Conv-TasNet

You need to modify the default parameters in the test_tasnet.py file, including test files, test models, etc.

For multi-audio

python test_tasnet.py 

For single-audio

python test_tasnet_wav.py 

Dual-Path-RNN

You need to modify the default parameters in the test_dualrnn.py file, including test files, test models, etc.

For multi-audio

python test_dualrnn.py 

For single-audio

python test_dualrnn_wav.py 

Pretrain Model

Conv-TasNet

Conv-TasNet model

Dual-Path-RNN

Dual-Path-RNN model

Result

Conv-TasNet

Final Results: 15.8690 is 0.56 higher than 15.3 in the paper.

Dual-Path-RNN

Final Results: 18.98 is 0.1 higher than 18.8 in the paper.

Reference

1. Luo Y, Chen Z, Yoshioka T. Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation[J]. arXiv preprint arXiv:1910.06379, 2019.
2. Conv-TasNet code && Dual-RNN code