pigip / ERNN-for-speech-enhancement

Equilibriated RNN for speech enhancement

In this repository, real-time speech enhancement method using the equilibriated recurrent neural network (ERNN) for the T-F mask estimator [1] is impremented using Chainer.

Our paper can be found here. If you use codes in this repository, please cite this paper.

In our paper, VoiceBank-DEMAND dataset [2] (available here) is used.

Dependencies

We have tested these codes on follwoing environment:

• Python 3.6.4
• Chainer 6.2.0
• NumPy 1.17.2
• CuPy 6.2.0
• CUDA Runtime Version 10100
• cuDNN Version 7500

Usage example

A set of Python codes for training and test are available.

Run "01_train.py" to train a model

Run "02_test.py" to evaluate a model

Note that paths in each code need to be changed for your environment.

Reference

[1] D. Takeuchi, K. Yatabe, Y. Koizumi, Y. Oikawa, and N. Harada, “Real-time speech enhancement using equilibriated RNN,” in 2020 IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2020. (accepted)

[2] C. Valentini-Botinho, X. Wang, S. Takaki, and J. Yamagishi, “Investigating RNN-based speech enhancement methods for noise-robust Text-to-Speech.,” in 9th ISCA Speech Synth. Workshop, 2016, pp. 146–152.