Multi-band spectral subtraction method combined with auditory masking properties for speech enhancement

Abstract

In order to reduce the music noise introduced by conventional spectral subtraction method for speech enhancement, we put forward a speech enhancement algorithm based on the combination of multi-band spectral subtraction and the masking properties of human auditory system. Firstly, we use the weighted recursive averaging method to estimate the noise power spectrum, exert subtraction of multi-band on the noise-corrupted speech signal; then, compute auditory masking threshold using the estimated speech signal, adjust the subtraction factor according to the threshold masking; finally, we obtain the spectrum of enhanced speech through computing the gain faction according to the subtraction factor. The simulation shows that, at low SNR, compared with conventional spectral subtraction, background noise and residual music noise have been effectively suppressed, and the clarity and intelligibility of speech signal have been obviously improved.