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

Bone conduction (BC) sensors naturally resist ambient noise, but the captured speech suffers from severe high-frequency attenuation due to the low-pass filtering characteristics of body tissue. To compensate for this hardware-induced information deficiency, we propose a time-domain framework leveraging highly generalized representations from Self-Supervised Learning (SSL). Specifically, we employ a large-scale pre-trained SSL model to generate embeddings that function as robust acoustic priors. Subsequently, a Key-Value Memory module is integrated to bridge the sensor domain gap, enabling the retrieval of high-fidelity priors from BC queries in the absence of reference air conduction signals. These retrieved cues are then processed by a Gated Attention Projection and dynamically fused into the primary network’s bottleneck, effectively recovering the high-frequency harmonics attenuated by the physical transmission path and rectifying the spectral distortion inherent in BC signals. Experiments on the ABCS and ESMB datasets demonstrate that our method surpasses state-of-the-art baselines in both quality and efficiency. It achieves PESQ gains of over 51% and 73% relative to raw BC inputs, respectively, with a compact architecture optimized for real-world deployment.