AN LP-BASED BLIND RESTORATION METHOD FOR IMPROVING INTELLIGIBILITY OFBONE-CONDUCTED SPEECH

摘要

Bone-conducted speech can be used instead of air-conduced speech in a noisy environment. However, its intelligibility is degraded when transmitted through bone-conduction. Therefore, blindly improving intelligibility of bone-conducted speech is a challenging topic. In previous study, we proposed an LP-based model which showed the expressed ability of blindly improving voice-quality of bone-conducted speech. In this report, we improve the LP-based model by (1) using LSF coefficients on LP representation and (2) using a recurrent neutral network for predicting parameters on frame by frame basis. We evaluated the proposed model in comparison with other models to know whether these can adequately improve intelligibility of BC speech signals and whether these can blindly work well. These evaluations were done with both subjective (MOS) and objective (LSD, MCD, and LCD) measurements. The experimental results of our proposed model showed the significant capability of improving the intelligibility of bone-conducted speech.