NEURAL CASCADE ARCHITECTURE FOR JOINT ACOUSTIC ECHO AND NOISE SUPPRESSION

Area of Science:

• Signal Processing
• Artificial Intelligence
• Speech Enhancement

Background:

• Acoustic echo and noise significantly degrade speech quality in communication systems.
• Existing methods often struggle with simultaneous suppression of both echo and noise while preserving speech fidelity.

Purpose of the Study:

• To propose a novel neural cascade architecture for joint acoustic echo and noise suppression.
• To enhance speech quality by effectively mitigating acoustic impairments.

Main Methods:

• A two-module neural cascade architecture was developed, featuring a Convolutional Recurrent Network (CRN) for complex spectral mapping and a Long Short-Term Memory (LSTM) network for magnitude mask estimation.
• The architecture was trained end-to-end with a unified loss function, optimizing both modules jointly.
• The final enhanced signal was reconstructed using phase information from the first module and magnitude information from the second module.

Main Results:

• The proposed cascade architecture demonstrated robust magnitude estimation and effective phase enhancement.
• Experimental results confirmed significant suppression of acoustic echo and noise.
• The method successfully preserved high speech quality, outperforming contemporary approaches.

Conclusions:

• The proposed neural cascade architecture offers an effective solution for joint acoustic echo and noise suppression.
• End-to-end training and the synergistic interaction between CRN and LSTM modules contribute to superior performance.
• This approach represents a significant advancement in speech enhancement technology.