Enhancing Bone Conduction Sensor Signals via Self-Supervised Acoustic Priors and Key-Value Memory.

Area of Science:

• Signal Processing
• Machine Learning
• Biomedical Engineering

Background:

• Bone conduction (BC) sensors offer noise resistance but suffer from high-frequency attenuation due to body tissue's low-pass filtering.
• This attenuation degrades speech signal quality, necessitating advanced compensation techniques.

Purpose of the Study:

• To develop a time-domain framework using Self-Supervised Learning (SSL) to restore high-frequency information in BC speech signals.
• To bridge the sensor domain gap and recover spectral distortion without reference air conduction signals.

Main Methods:

• Leveraging a large-scale pre-trained SSL model to generate robust acoustic priors from BC signals.
• Integrating a Key-Value Memory module for retrieving high-fidelity priors.
• Utilizing Gated Attention Projection for dynamic fusion and high-frequency harmonic recovery.

Main Results:

• The proposed method significantly improves speech quality, achieving over 51% and 73% PESQ gains on the ABCS and ESMB datasets, respectively.
• Demonstrated superior performance compared to state-of-the-art baselines in both quality and efficiency.
• The compact architecture is optimized for practical, real-world deployment.

Conclusions:

• The SSL-based framework effectively compensates for hardware-induced deficiencies in BC sensors.
• This approach offers a promising solution for enhancing BC speech intelligibility and quality.
• The method presents an efficient and effective way to rectify spectral distortion in BC signals.