Bandwidth Extension System for Throat Microphone Speech Reconstruction

The throat microphone is effective for capturing speech in noisy environments but typically produces narrowband signals, requiring post-processing for speech enhancement. Additionally, the limited availability of throat speech (TS) datasets poses challenges for system development. To address this, we introduce Dent-DDSP, a robust speech generator that simulates TS using existing air-conducted speech (AS) datasets. We employ bandwidth extension (BWE) methods to recover missing high-frequency components and use TUNet, a DNN-based model combining transformer and UNet structures. Experimental results demonstrate the superior speech quality enhancement and the unit language model’s better perceptual performance, highlighting promising avenues for throat microphone applications.