Adaptive time-frequency data fusion for speech enhancement

This paper proposes an adaptive time- frequency data fusion technique for the reduction of noise in speech signals using an array of two mi- crophones. Recently, it has been shown that phase- error filtering can be a simple and effective method for "cocktail party" noise removal. However, while this technique is successful when the recorded speech sig- nal is veq noisy (less than lOdB signal-to-noise ratio (SNR)), it also tends to severely degrade the signal at higher SNRs. In this paper, the phase-error filtering technique is extended to dynamically estimate the SNR and adjust the filter parameters accordingly. Simula- tion results show an SNR gain of lldB (with Gaussian noise) and an SNR gain of 17dB (Speech noise) at low SNRs without any signal degradation at higher SNRs. Speaker-independent speech recognition results using 5 speakers show that the proposed algorithm achieves a digit percent accuracy gain of 22% at OdB and 15% at 2OdB.