RemixDPT-EGNet: remix-based bone-conducted speech enhancement in the time domain

Due to the complexity of the bone-conducted (BC) speech application environment, obtaining aligned air-conducted (AC) - BC speech data for model training still faces great limitations. We propose RemixDPT-EGNet, a novel self-supervised BC speech enhancement model that aims to reconstruct high-quality AC speech directly from BC speech input without relying on the corresponding clean AC speech during inference. It extends the RemixIT architecture, adopts a task-specific teacher-student model design in the time domain, and is optimized for low-resource and noisy environments. It introduces a transformation function that extracts individual characteristics between BC and AC speech, enabling the model to fit more BC speech for data enhancement and better enhance the generalization ability of the model. Results show that this method significantly outperforms with a PESQ improvement of 51.8% compared to the input BC speech, and a better STOI score. The study also confirmed the key impact of BC microphone position, among which the throat and ear positions can bring the most reliable enhancement effect. In addition, the transformation function significantly improves the various evaluation values of the output speech, performs well under heavy compression, and maintains clarity even with very little spectral information.