Fast HMM-driven beamforming for robust speech recognition in reverberant environments

The reverberation is induced from the combination of the direct waveform and multiple reflected waveforms. Therefore, the reverberation is highly corrected with the original speech signal. This leads to dramatically degrade the performance of speech recognition. This paper extends VTS methodology to develop a robust technique for HMM-driven beamformer on the reverberation environments. We approximate the logarithm of Gaussian mixture models of HMM with VTS expansion. This makes it possible to obtain a simpler updating functions of beamformer parameters than the counterparts of original HMM-driven beamformer. The RWCP database is used for the simulation of the multi-channel recorded reverberation speech. A speaker-independent speech query task of Mandarin names was applied to evaluate the performance of the beamformers. Our experimental results show that the proposed algorithm was effective on compoutation reduction for the adaptation process of HMM-driven beamformer. It indicates that the proposed framework benefits the development in robust speech recognition on resource-constrained platforms.