Dereverberation based on Wavelet Packet Filtering for Robust Automatic Speech Recognition

This paper describes a multiple-resolution signal analysis to suppress late reflection of reverberation for robust automatic speech recognition (ASR). Wavelet packet tree (WPT) decomposition offers a finer resolution to discriminate the late reflection subspace from the speech subspace. By selecting appropriate wavelet basis in the WPT for speech and late reflection, we can effectively estimate the Wiener gain directly from the observed reverberant data. Moreover, the selection procedure is performed in accordance with the likelihood of acoustic model used by the speech recognizer. Dereverberation is realized by filtering the wavelet packet coefficients with the Wiener gain to suppress the effects of the late reflection. Experimental evaluations with large vocabulary continuous speech recognition (LVCSR) in real reverberant conditions show that the proposed method outperforms conventional wavelet-based methods and other dereverberation techniques.