Computationally Efficient Body-Conducted Voice Conversion with Original Excitation Signals

In this paper, we propose a computationally efficient method of body-conducted voice conversion. A body-conducted voice is robust against to external noise but its voice quality is severely degraded by mechanisms of body-conduction. The conventional body-conducted voice conversion method effectively enhances the body-conducted voice by converting both spectral and excitation features. On the other hand, its computational cost is relatively high. To significantly reduce the computational cost while keeping the enhanced voice quality as high as possible, we propose a conversion method of using an original excitation signal of the body-conducted voice and computationally efficient feature extraction. The effectiveness of the proposed method is confirmed in the objective and subjective evaluations.