A Pytorch implementation of Google's VoiceFilter System

1. Data simulation

   ./nnet/data_simulate.py --dump-dir simu/train /path/to/librispeech/train.scp asset/train_tuples.csv
   ./nnet/data_simulate.py --dump-dir simu/dev /path/to/librispeech/dev.scp asset/dev_tuples.csv

2. Speaker embedding (I used public xvector from here)

3. Data prepare

   Prepare data as {mix,ref,emb}.scp and the format of scp file follows Kaldi's recipe, egs <key> <path> pair in each line.

4. Confugure nnet/conf.py and train the model (see train.sh for details).

5. Use nnet/separate.py for inference.

1. I used Si-SNR loss instead of MSE of spectrogram, which could achieve better perfermance.