Separation of mixed speech signals of short duration using Wiener filter as postprocessing for Frequency Domain ICA

AKIHIRO HANDA; LEANDRO DI PERSIA; KENKO OHTA; MASUZO YANAGIDA

Atami

Congreso; Monthly meeting of the Special Interest Group on Spoken Language Processing; 2006

Information Processing Society, Japan

Frequency-domain ICA is effective for separating mixed speech signals of long duration but it is not the case for signals of short duration in environments having ordinary reverberation time. The main reason would be lack of data in each frequency bin. The optimal window size and shifting interval for separating short speech are investigated, and Wiener filter is adopted in each frequency bin as post-processing of frequency-domain ICA.