Adaptive spectral subtraction for robust speech recognition

초록

Speech recognition rate degrades drastically in extreme noisy environments. Spectral subtraction is one of the representative noise reduction method, but it is vulnerable to non-stationary noise although it is quite effective for stationary noise. In this paper, we propose an adaptive spectral subtraction method to improve the speech recognition performance. The proposed method is to consistently update the noise component in non-speech regions and remove the corresponding component in following speech regions. To validate of the noise reduction performance, we conducted several experiments for each noise power level. Our approach achieved better performance compared to the conventional spectral subtraction approach.

키워드