A Noise Suppression Method Based on Spectral Subtraction and RASTA for Speech Recognition

Discusses noise suppression in speech recognition under conditions of additive and convolutional noise. In the step of feature extraction the effect of additive noise is compensated by spectral subtraction based on short time mean values of the power spectrum, and convolutional noise is compensated by the RASTA (relative spectral) technology on Mel frequency cepstrum. Experiments on speaker independent mandarin isolated digits recognition showed that word error rates with the proposed method are lower than those without it. Moreover, the method needs less apriori knowledge of the noise and has lower complexity in calculation. The proposed method based on spectral subtraction and RASTA is an effective method in noise suppression.