Blind Signal Separation and Speech Recognition in the Frequency Domain

In this paper it is shown that a Blind Signal Separation (BSS) method in the frequency domain (FDBSS) improves significantly the speaker Signal to Interference Ratio (SIR) and the phoneme recognition score of a continuous speech, speaker-independent acoustic decoder in a multi-simultaneous-speaker office environment. Specifically, the efficiency of the presented FDBSS method is studied on a TITO (Two Input, Two Output) network. In extensive experiments in an artificially created environment using real-room impulse responses, the mean SIR resulting from the Output Decorrelation was increased by approximately 8dB. Furthermore, a percentage phoneme recognition improvement of 85% and 116% for each one of the separated speech signals compared to the mixed signals was measured. It is also shown that the complexity of the FDBSS method is significantly lower than in the time domain and for M-order linear separating filters is O(MlogM) compared to the O(M) in the time domain.