Robust Speech Recognition in Noisy and Reverberant Environments Using Wavelet-based Wiener Filtering
Randy Gómez, Tatsuya Kawahara · 2011
We present a method of enhancing the speech signal corrupted by noise and late reflection in the wavelet domain for robust automatic speech recognition (ASR). The wavelet parameters for speech, background noise and late reflection are optimized to achieve a better estimate of the Wiener gain for effective filtering. Wiener gains to compensate for the effects of background noise and late reflection are independently estimated and then combined. To cope with different noise and reverberant conditions, we introduce the noise profiles and reverberation time identification. The proposed method is evaluated in a large vocabulary continuous speech recognition (LVCSR) task, and shown to outperform several conventional methods.