Semi-Supervised Enhancement and Suppression of Self-Produced Speech Using Correspondence between Air- and Body-Conducted Signals

We propose a semi-supervised method for enhancing and suppressing self-produced speech recorded with wearable air- and body-conductive microphones. Body-conducted signals are robust against external noise and predominantly contain self-produced speech. As a result, these signals provide informative acoustical clues when estimating a linear filter to separate a mixed signal into self-produced speech and background noise. In a previous study, we proposed a blind source separation method for handling air- and body-conducted signals as a multi-channel signal. While our previously proposed method demonstrated the superior performance that can be achieved by using air- and body-conducted signals in comparison to using only air-conducted signals, the enhanced and suppressed air-conducted signals tended to be contaminated with the acoustical characteristics of the body-conducted signals due to the nonlinear relationship between these signals. To address this issue, in this paper, we introduce a new source model which takes into consideration the correspondence between these signals and incorporates them within a semi-supervised framework. Our experimental results reveal that this new method alleviates the negative effects of using the acoustical characteristics of the body-conducted signals, outperforming our previously proposed method, as well as conventional methods, under a semi-supervised condition.