NOISE SURPRESSION

v1.0.0

Python package for noise supression in audio based on DNN

Acknowledgements

Based on the Yong Xu and Qiuqiang Kong SEDDN package https://github.com/yongxuUSTC/sednn

Model

A trained model (https://github.com/nsu-ai-team/noise_supression/blob/master/model/md_10000iters.h5) is distributed with the python package.

Model information:

1. Trained on a subset of the LibriSpeech audio corpus (14.6 hours)
2. PESQ=1.82 +- 0.24
3. Trained using the sednn/mixture2clean_dnn subrepo
4. Audio with english speech are used, but can show relatively good quality on other languages
5. Will show worse enhancement results on data from other corpuses (or real-life data), keep that in mind

Installation

pip install git+https://git@github.com/nsu-ai-team/noise_supression.git

Python usage

Loading the module

>>> from noise_supression.enhance_audio import AudioEnhancer
>>> ae = AudioEnhancer()

Enhancing an audio file

>>> ae.enhance_audio('example/61-70968-0000.n.wav','example/61-70968-0000.enh.wav')

Command line usage (if cloned)

Loading the module

$ python enhance_audio.py example/61-70968-0000.n.wav example/61-70968-0000.enh.wav