Speech enhancement and improved recognition accuracy by integrating wavelet transform and spectral subtraction algorithm

Spectral subtraction (SS) approach has been widely used for speech enhancement and recognition accuracy improvement, but becomes less effective when the additive noise is not white. In this paper, we propose to integrate wavelet transform and the SS algorithm. The spectrum of the additive noise in each frequency band obtained in this way can then be better approximated as white if the number of bands is large enough, and therefore the SS approach can be more effective. Experimental results based on three objective performance measures and spectrogram-plot comparison show that this new approach can provide better performance especially when the noise is non-white. Listening test results also indicate that the new algorithm can give more preferable sound quality and intelligibility than the conventional spectral subtraction algorithm. Moreover, the new approach also offers some reductions of the computational complexity when compared with the conventional SS algorithm.