Multi-TALK: Multi-Microphone Cross-Tower Network for Jointly Suppressing Acoustic Echo and Background Noise

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Acoustic echo and background noise significantly degrade audio quality in multi-channel communication systems.
• Existing noise suppression methods often struggle with balancing noise reduction and speech distortion.

Purpose of the Study:

• To propose and evaluate a novel multi-channel network, Multi-TALK, for simultaneous acoustic echo and background noise suppression.
• To enhance the performance of noise suppression algorithms by minimizing speech distortion.

Main Methods:

• Developed a multi-channel cross-tower network with attention mechanisms in the latent domain (Multi-TALK).
• Employed a parallel encoder with an auxiliary encoder for multi-channel feature extraction and spatial information compression.
• Utilized attention mechanisms to integrate far-end speech features and iteratively refine noise estimation within the cross-tower network.
• Incorporated attention-based post-processing to mitigate over-suppression and preserve near-end speech quality.

Main Results:

• The proposed Multi-TALK algorithm effectively suppresses both acoustic echo and background noise.
• Demonstrated a significant reduction in speech distortion compared to conventional algorithms.
• Achieved improved audio clarity through advanced latent domain processing and attention mechanisms.

Conclusions:

• Multi-TALK offers a robust solution for acoustic echo and background noise suppression in multi-channel scenarios.
• The attention mechanisms and iterative estimation within the latent domain are key to the algorithm's effectiveness.
• The approach successfully balances noise reduction with the preservation of near-end speech integrity.