Array configuration-agnostic personalized speech enhancement using long-short-term spatial coherence.

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Personalized speech enhancement (PSE) aims to suppress interfering speech.
• Multichannel PSE systems offer advantages over single-channel methods in noisy environments.
• Implementing multichannel PSE for diverse household microphone arrays is challenging.

Purpose of the Study:

• To develop an array configuration-agnostic multichannel personalized speech enhancement (PSE) system.
• To introduce a novel spatial feature for monitoring target speaker activity.
• To evaluate the system's performance in adverse acoustic conditions.

Main Methods:

• A novel spatial feature, long-short-term spatial coherence (LSTSC) with a dynamic forgetting factor, was defined.
• A convolutional recurrent network was utilized with the LSTSC feature as input.
• A simplified LSTSC feature was proposed to reduce computational cost.
• Experiments compared the proposed PSE systems against baselines using unseen room responses and array configurations.

Main Results:

• The proposed multichannel PSE network trained with LSTSC achieved superior speech enhancement.
• The system demonstrated effectiveness without requiring precise knowledge of array configurations or room responses.
• Both complete and simplified LSTSC versions showed significant performance improvements.

Conclusions:

• The developed LSTSC feature enables array configuration-agnostic multichannel PSE.
• The proposed system offers a robust solution for speech enhancement in complex acoustic environments.
• This approach advances the practical application of multichannel PSE in real-world scenarios.