This is an old implementation for "Multi-Domain Processing via Hybrid Denoising Networks for Speech Enhancement" (2018, arxiv, demo).

Dependencies

This code was written with tensorflow <= 1.4.0 and python 3.5

Contents

• speech data pre-processing codes (tfrecordreadwrite.py)
• differentiable STFT module (Utils.py)
• Various speech enhancement models (net.py: U-Net, LSTM on Spectrogram, TasNet on waveform)
• Hybrid models (model.py)
• Various objective functions for speech enhancement (loss.py)

To run codes,

python main.py

Arguments are listed in configuration.py.

Results

Note, 'Ours' represents a multi-domain hybrid model.