Direction-Based Speech Enhancement Using Multichannel Subband-Fullband Neural Network

Here are some demos on the proposed direction-based multichannel subband-fullband method. Find the paper here: (https://doi.org/10.1109/WASPAA58266.2023.10248181)

Audio and video is taken from the Easycom dataset by Donley et al. (https://doi.org/10.48550/arXiv.2107.04174).

Note that the Maximum-Directivity MVDR baseline method shown here is suboptimal for multi-speaker enhancement, which has therefore been disabled for this method.

This demo page is authored by Benjamin Stahl, IEM Graz, 2023

Synthetic Example (Overlapping Speech)

Mixture Speaker1-MVDR Speaker2-MVDR Speaker1-Proposed Speaker2-Proposed