Cepstral smoothing of masks for single-channel speech segregation

In this paper, cepstral smoothing is introduced to reduce musical noise in speech separated using computational auditory scene analysis (CASA). Our post-processing algorithm is composed of three steps. First, speech and interference are separated from noisy speech. Next, binary masks in the time-frequency domain are obtained using separated speech and noise signals. Finally, a mask smoothing algorithm which preserves speech onsets and quasi-stationary narrow-band structures is used to eliminate excessive peaks in time-frequency domain that lead to musical noise. Using an objective speech quality measure, we show that proposed method can improve speech quality.