Optimizing Deep Learning for Joint Blind Source Separation and Noise Suppression in Audio

This study introduces a deep learning-based network that jointly addresses blind source separation and noise suppression in audio signals, overcoming limitations of conventional methods that treat these problems separately. The network integrates multiple components including a learnable encoder and a multi-scale dilated separator with attention mechanisms, trained end-to-end with a three-term objective function. Evaluations utilize datasets such as LibriMix, WHAM!, and WHAMR! to assess performance under challenging noise conditions.

Originally published by gnews