A Speech Separation Method Combining Time-Frequency Masking and Independent Component Analysis

The anechoic mixing model, the number of sources is larger than the number of mixtures, and sources have different amplifications and time delays in different mixtures, was effectively solved by the speech separation method based on time-frequency masking. As the binary winner-take-all masks were created to extract the sources, the constructed masks contain a large number of zeros, and bring in music noise or result in speech distortion. Meanwhile, in some applications, there is no need to extract all the sources. Considering these two issues above, a speech separation method combining time-frequency masking and independent component analysis (ICA) is proposed in this paper. The time-frequency masking technique is used to remove the sources rather than extract them. The ICA algorithm is exploited to separate the needed sources that are useful to us. Experimental results show the effectiveness of the proposed method.