Deep Learning Based Target Cancellation for Speech Dereverberation.

Area of Science:

• Speech processing
• Deep learning
• Acoustic signal processing

Background:

• Reverberation significantly degrades speech quality and intelligibility.
• Existing dereverberation methods, like weighted prediction error (WPE), have limitations in complex acoustic environments.

Purpose of the Study:

• To develop and evaluate deep learning models for single- and multi-channel speech dereverberation.
• To enhance speech recognition performance in reverberant conditions.

Main Methods:

• Single-channel: Extended magnitude-domain processing to complex-domain mapping using deep neural networks (DNNs) to predict direct-path signal components.
• Multi-channel: Employed a minimum variance distortionless response (MVDR) beamformer and fed its output (RI components) as features to DNNs for dereverberation.

Main Results:

• The proposed DNN-based models achieved excellent speech dereverberation and recognition performance.
• Consistently outperformed single- and multi-channel weighted prediction error (WPE) algorithms on the REVERB challenge test set.

Conclusions:

• Deep learning offers a powerful approach for advanced speech dereverberation.
• The proposed complex-domain mapping and multi-channel integration significantly improve performance over traditional methods.