Quantile based speech/noise-dominant decision for wavelet based speech enhancement

Summary form only given. This paper presents a quantile based speech/noise-dominant decision for wavelet based speech enhancement. In the proposed decision, the decision between the speech dominant and noise one is carried out using temporal quantiles for previous frames, and the influence of noise is reduced by subtracting the noise component. According to the decision, the speech component is preserved in the speech dominant, and noise is fully reduced in the noise dominant. Therefore, the enhanced speech reduces noise and speech distortion. Moreover, the proposed method also requires no speech activity detector or signal statistics. From the investigation of the Itakura-Saito distortion measure, segmental SNR and listening tests, we show the proposed method is more effective than the conventional method in reducing noise without introducing speech distortion.