A Convolutional Recurrent Neural Network for Real-Time Speech Enhancement

Implement A Convolutional Recurrent Neural Network for Real-Time Speech Enhancement by PyTorch.

Performance

Model performance on private dataset, for reference only

Experiment		5dB			10dB		Average	Comment
	0.5m	1m	2m	0.5m	1m	2m		Distance to the microphone
Mixture	2.33	2.17	1.85	2.44	2.27	1.94	2.167
LSTM	2.62	2.49	2.02	2.71	2.53	2.13	2.417
Our implementation	2.630	2.458	2.086	2.729	2.527	2.172	2.434
Our implementation (LN)	2.703	2.461	1.961	2.796	2.548	2.181	2.442	Replace all batch norm with layer norm

Dependencies

• Python3
• torch==1.1.0
• librosa==0.7.0
• SoundFile==0.10.2
• tensorboard==1.14.0
• tensorboard==1.13.1(for visualization only)
• pypesq==1.0, pip install https://github.com/vBaiCai/python-pesq/archive/master.zip
• pystoi==0.2.2
• matplotlib==3.1.0
• tqdm==4.32.2

Reference

• CRNN_mapping_baseline
• A Convolutional Recurrent Neural Network for Real-Time Speech Enhancement
• EHNet
• Convolutional-Recurrent Neural Networks for Speech Enhancement