A Modulation-Domain Loss for Neural-Network-Based Real-Time Speech Enhancement - ICASSP 2021

Official repo for the PyTorch implementation of the spectro-temporal modulation-domain loss presented at ICASSP 2021

Modulation loss function is used for training DNN-based speech enhancement systems
Can combine with any other loss function

[Arxiv] [Paper]

MAKE SURE THE KERNEL WEIGHTS ARE NOT BEING UPDATED FROM BACKPROPAGATION

See example usage in modulation_loss.py

Gabor-STRF Kernel Examples

A MODULATION-DOMAIN LOSS FOR NEURAL-NETWORK-BASED REAL-TIME SPEECH ENHANCEMENT

Citation

@INPROCEEDINGS{vuong2021modulation,
  author={Vuong, Tyler and Xia, Yangyang and Stern, Richard M.},
  booktitle={ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={A Modulation-Domain Loss for Neural-Network-Based Real-Time Speech Enhancement}, 
  year={2021},
  volume={},
  number={},
  pages={6643-6647},
  doi={10.1109/ICASSP39728.2021.9414965}}