Speech separation using DUET and binary masking with temporal smoothing in cepstral domain

This paper presents a blind speech separation (BSS) method based on Degenerate Unmixing Estimation Technique (DUET) and binary time-frequency masking approach. Firstly, the speech signals are extracts from the anechoic speech mixtures using the DUET algorithm. Then, a temporal smoothing procedure is applied in the cepstral domain to the spectral masks. The latter are estimated from the obtained signals. The smoothing procedure is capable of reducing musical noise artifacts in the separated signal which is often generated by the time-frequency binary mask based BSS approaches. Experimental results indicate that the proposed method provides high interference suppression and musical noise reduction.