DANet-For-Speech-Separation

January 9, 2020 ยท View on GitHub

Pytorch implement of DANet For Speech Separation

Chen Z, Luo Y, Mesgarani N. Deep attractor network for single-microphone speaker separation[C]//2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017: 246-250.

Requirement

  • Pytorch 0.4.0
  • librosa 0.7.1
  • PyYAML 5.1.2

Due to the RNN multi-GPU parallel problem, only PyTorch 0.4.0 is supported.

Training steps

  1. First, you can use the create_scp script to generate training and test data scp files.
python create_scp.py
  1. Then, in order to reduce the mismatch of training and test environments. Therefore, you need to run the util script to generate a feature normalization file (CMVN).
python ./utils/util.py
  1. Finally, use the following command to train the network.
python train.py -opt ./option/train.yml

The repository model code draws on DANet from naplab. It is found through experiments that the loss value cannot be decreased, and I have limited ability and have not identified the cause of the problem. Therefore, only training codes are released for reference only.