Implementation of Speech Separation Model

This is a Tensorflow implementation of Speaker Indepentent Source Separation.

Models we have implemented are (1) TasNet and (2) Cross Domain Joint Embedding and Clustering Network.

Also, we have implemented an alternative to solve the label ambiguity problem, described in Interrupted and Cascaded PIT.

Environments

You can run on TensorFlow 2 !!! However, instead of executing eagerly, we build the graph first as done in TF v1.

Results

No.	Model	Label Assignment	SDRi (Validation)	SDRi (Test)
(1)	Tasnet	PIT	16.2 dB	15.8 dB
(2)	CDNet	PIT	17.1 dB	16.9 dB
(3)	TasNet	Fixed Assign (L=100)	17.3 dB	16.9 dB
(4)	TasNet	Fixed Assign (L=80)	17.7 dB	17.4 dB
(5)	TasNet Init from (4)	PIT	18.0 dB	17.7 dB

Usage

Training:

python main.py -m train -c json/config.json

Testing:

python main.py -m test -c models/name/config.json -ckpt chosen_checkpoint

IAC PIT Training:

1st stage :

Train a model using PIT first : python main.py -m train -c json/tasnet-1.json

2nd stage :

Extract the pretrained label assignment of a 1st stage model by using write_pretrained_perm('tasnet-1', 100) in util.py to generate a fixed label assignment file in models/tasnet-1/perm_idx/100.csv (default epoch = 100)

Then train the model using python main.py -m train -c json/tasnet-2.json

3rd stage :

Load the 2nd stage model parameters, and continue to train the model with PIT.

Which can be simply done by python main.py -m train -c json/tasnet-3.json.

Configuration

A detailed description of all configurable parameters can be found in json/tasnet-1.json

Optional command-line arguments:

Argument	Valid Inputs	Default	Description
mode	train/test	training
config	string	config.json	Path to JSON-formatted config file
ckpt	string	None	Path to model's checkpoint. If not specfied, will automatically load the latest checkpoint.

Dataset Preprocess

From SPHERE to wav : bash convert_wsj0.sh

Generate WSJ0-2mix (Wall Street Journal with 2-speaker mixture) or WSJ0-3mix

1. Download official code or use my modified version in create_wav_2speakers.m and create_wav_3speakers.m

2. Download voicebox

3. Steps to run octave on linux:

   (1) run octave-cli

   (2) load package pkg load <pkg-name>

   (3) run create_wav_2speakers.m or create_wav_3speakers.m

Reference

If you find this repo interesting, you can refer to more details in the following papers.

[1] Yang, G., Tuan, C., Lee, H., Lee, L. (2019) Improved Speech Separation with Time-and-Frequency Cross-Domain Joint Embedding and Clustering. Proc. Interspeech 2019, 1363-1367, DOI: 10.21437/Interspeech.2019-2181. Link to paper

[2] G. Yang, S. Wu, Y. Mao, H. Lee and L. Lee, "Interrupted and Cascaded Permutation Invariant Training for Speech Separation," ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, Spain, 2020, pp. 6369-6373, doi: 10.1109/ICASSP40776.2020.9053697. Link to paper