Spatial speech detection for binaural hearing aids using deep phoneme classifiers

Area of Science:

• Audiology and Speech Processing
• Signal Processing
• Artificial Intelligence in Hearing Aids

Background:

• Current hearing aids struggle with speech-specific optimization for spatial sound sources.
• Effective speech enhancement in noisy environments remains a challenge for binaural hearing aid users.

Purpose of the Study:

• To develop and evaluate a novel approach for spatial speech detection and blind speech enhancement in binaural hearing aids.
• To combine direction of arrival (DOA) estimation with automatic speech recognition-based speech quality measures for improved hearing aid performance.

Main Methods:

• Integrated a high-resolution DOA estimator with a low-resolution speech quality measure based on phoneme representations from a deep neural network (DNN).
• Explored three automatic speech quality measures (ASQM): entropy, mean temporal distance (M-Measure), and matched phoneme (MaP) filtering.
• Tested the approach in diverse acoustic scenes with varying signal-to-noise ratios (SNR), noise types (localized/diffuse), and reverberation levels.

Main Results:

• Two ASQMs, M-Measure and MaP filtering, reliably identified the speech target across different acoustic conditions.
• The proposed system demonstrated robust performance without environment adaptation or prior scene information.
• The approach successfully detected incorrect spatial filtering angles, indicating its effectiveness in real-world scenarios.

Conclusions:

• The developed approach enhances spatial speech detection and blind speech enhancement for binaural hearing aids.
• ASQMs based on M-Measure and MaP filtering show significant promise for improving hearing aid functionality in complex acoustic environments.
• The system's adaptability and lack of reliance on a priori information make it a practical solution for advanced hearing aid technology.