Spectral Subtraction Based on Speech/noise-dominant Classification

This paper presents a spectral subtraction using the classifications between the speech dominant and the noise one. In our system, a new classification scheme between the speech dominant and the noise one is proposed. The proposed classifications use the standard deviation of the spectrum of observation signal in each critical band. We have introduced two oversubtraction factors for speech dominant and noise one, respectively. And spectral subtraction is carried out after the classification. The proposed method is tested on several noise types from the Noisex-92 database. On the basis of segmental SNR, inspection of spectrograms and listening tests, the proposed system is shown to be effective to reduce background noise. Moreover, our system generates less musical noise and distortion than the conventional systems.