Parameter efficient speaker adaptation for enhancement of bone-conducted speech

Bone conducted speech (BCS) signals have reduced intelligibility and quality due to a limited frequency response and speaker-specific distortion introduced by bone conduction pathways. However, BCS is isolated from the environment, which may offer noise-free communication in high noise environments. Deep neural network speech models were designed to investigate the enhancement of BCS for new unseen speakers. Personalization to the new speakers was examined through full-model fine tuning and parameter-efficient adaptation. Listener subjective quality ratings and objective metrics of intelligibility and quality demonstrate significant enhancement of BCS for unseen speakers, with personalization using data-limited parameter-efficient model adaptation.