A consideration on time-frequency masking methods for speech separation

Time-Frequency Masking methods, primary known as DUET [2] and SAFIA [3], are effective scheme for blind speech separation problem. Based on an investigation of conventional delay-histogram and the time-frequency masking method in terms of estimated delay accuracy, two novel approaches for clustering process are proposed. In particular, the proposed methods tend to improve relatively large amount of delay estimation error using STFT phase difference in lower frequency band. A novel clustering idea is the use of frequency vs. phase difference dot plot in a frame by frame manner. The other one is to use filter bank and fractional delay operation. These approaches are proved to be effective through several experiments.