suwoncjh / TasNet-tensorflow

A tensorflow implementation of TasNet (ICASSP 2018)

TasNet: Time-domain Audio Separation Network

This is the implementation in Tensorflow of "TasNet: Time-domain Audio Separation Network for Real-time, single-channel speech separation", published in ICASSP2018, by Yi Luo and Nima Mesgarani.

Special

This implementation takes ododoyo's as reference, especially in SI-SNR and PIT training part. A extra MSE training objective and PIT training policy is implemented by myself. Also, this implementation haven't supported variable-length segments in training so far.
Discussion, (friendly) criticism, suggestions are always welcomed!

Requirements

• tensorflow 1.8.0
• python 3.5
• librosa

Contents

• params.py defines all global parameters.
• data_generator.py This file establishes WSJ0 2-mix datasets (referred to ICASSP 2016 Deep Clustering paper) and generates batch data for training. You may run this code firstly to generate datasets and change the path in tf_train.py.
• tf_net.py defines the TasNet structure, loss, training optimizer, etc.
• tf_train.py trains the model. Rewrite the dataset path with your own path.
• tf_test.py evaluates the model performance. This code hasn't been written well, still under repair.
• mir_eval.py and mir_util.py are forked from ododoyo's, implementing bss_eval calculation in Python rather than MATLAB.