Deep clustering for single-channel speech separation

Implement of "Deep Clustering Discriminative Embeddings for Segmentation and Separation"

Requirements

see requirements.txt

Usage

1. Configure experiments in .yaml files, for example: train.yaml

2. Training:

   python ./train_dcnet.py --config conf/train.yaml --num-epoches 20 > train.log 2>&1 &

3. Inference:

   python ./separate.py --num-spks 2 $mdl_dir/train.yaml $mdl_dir/final.pkl egs.scp
   

Experiments

Configure	Epoch	FM	FF	MM	FF/MM	AVG
config-1	25	11.42	6.85	7.88	7.36	9.54

Q & A

1. The format of the .scp file?

   The format of the wav.scp file follows the definition in kaldi toolkit. Each line contains a key value pair, where key is a unique string to index audio file and the value is the path of the file. For example

   mix-utt-00001 /home/data/train/mix-utt-00001.wav
   ...
   mix-utt-XXXXX /home/data/train/mix-utt-XXXXX.wav
   

2. How to prepare training dataset?

   Original paper use MATLAB scripts from create-speaker-mixtures.zip to simulate two- and three-speaker dataset. You can use you own data source (egs: Librispeech, TIMIT) and create mixtures, keeping clean sources at meanwhile.

Reference

1. Hershey J R, Chen Z, Le Roux J, et al. Deep clustering: Discriminative embeddings for segmentation and separation[C]//Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE International Conference on. IEEE, 2016: 31-35.
2. Isik Y, Roux J L, Chen Z, et al. Single-channel multi-speaker separation using deep clustering[J]. arXiv preprint arXiv:1607.02173, 2016.