Parameter efficient speaker adaptation for enhancement of bone-conducted speech

Area of Science:

• Speech processing
• Biomedical engineering
• Machine learning

Background:

• Bone-conducted speech (BCS) signals suffer from limited frequency response and speaker distortion, reducing intelligibility.
• Despite quality issues, BCS offers noise-free communication potential in loud environments due to its isolation from ambient noise.

Purpose of the Study:

• To investigate deep neural network (DNN) models for enhancing BCS quality and intelligibility.
• To explore personalization methods for DNN models to adapt to new, unseen speakers for BCS enhancement.
• To evaluate the effectiveness of full-model fine tuning versus parameter-efficient adaptation for personalization.

Main Methods:

• Developed DNN speech models specifically for BCS enhancement.
• Implemented and compared two personalization strategies: full-model fine tuning and parameter-efficient adaptation.
• Evaluated performance using listener subjective quality ratings and objective intelligibility/quality metrics.

Main Results:

• DNN models significantly enhanced the intelligibility and quality of BCS for unseen speakers.
• Parameter-efficient adaptation demonstrated effective personalization, even with limited speaker-specific data.
• Both subjective and objective evaluations confirmed the substantial improvements in BCS performance.

Conclusions:

• Deep neural networks, particularly with parameter-efficient adaptation, can effectively enhance bone-conducted speech for new speakers.
• Personalization is crucial for optimizing BCS enhancement models.
• This approach holds promise for improving communication in high-noise environments using BCS technology.