Virtual Acoustic Channel Expansion Based on Neural Networks for Weighted Prediction Error-Based Speech Dereverberation

Joon-Young Yang, Joon-Hyuk Chang


In this study, we propose a neural-network-based virtual acoustic channel expansion (VACE) framework for weighted prediction error (WPE)-based speech dereverberation. Specifically, for the situation in which only a single microphone observation is available, we aim to build a neural network capable of generating a virtual signal that can be exploited as the secondary input for the dual-channel WPE algorithm, thus making its dereverberation performance superior to the single-channel WPE. To implement the VACE-WPE, the neural network for the VACE is initialized and integrated to the pre-trained neural WPE algorithm. The entire system is then trained in a supervised manner to output a dereverberated signal that is close to the oracle early arriving speech. Experimental results show that the proposed VACE-WPE method outperforms the single-channel WPE in a real room impulse response shortening task.


 DOI: 10.21437/Interspeech.2020-1553

Cite as: Yang, J., Chang, J. (2020) Virtual Acoustic Channel Expansion Based on Neural Networks for Weighted Prediction Error-Based Speech Dereverberation. Proc. Interspeech 2020, 3930-3934, DOI: 10.21437/Interspeech.2020-1553.


@inproceedings{Yang2020,
  author={Joon-Young Yang and Joon-Hyuk Chang},
  title={{Virtual Acoustic Channel Expansion Based on Neural Networks for Weighted Prediction Error-Based Speech Dereverberation}},
  year=2020,
  booktitle={Proc. Interspeech 2020},
  pages={3930--3934},
  doi={10.21437/Interspeech.2020-1553},
  url={http://dx.doi.org/10.21437/Interspeech.2020-1553}
}