Robust Speech Recognition Using Optimized Wavelet Filtering in Reverberant Conditions

Speech recognition in reverberant environments is a difficult task. Reverberation has the effect of degradation of recognition performance due to acoustic mismatch. We present an optimization method of the wavelet parameters for dereverberation in automatic speech recognition (ASR). By tuning the wavelet parameters to improve the acoustic model likelihood, waveletbased dereverberation methods become more effective in the ASR application. We evaluate several existing wavelet-based methods and optimize them, based on our proposed scheme. Experimental evaluations through ASR experiments demonstrate significant improvement for all methods with the proposed optimization.