Perceptual wavelet filtering for robust speech recognition

In this paper an enhanced noise reduction for robust speech recognition is implemented by means of a perceptual wavelet filtering algorithm. The psychoacoustic model is applied to map the universal thresholds to the perceptually universal thresholds for each critical wavelet subband. By improving our quantile filtering method, the change of noise level is tracked more adaptively. The denoising algorithm is compared with well-known noise reduction methods embedded in different state-of-the-art speech recognizers. We achieve almost similar recognition performance with the HTK recognizer on AURORA3 SPEECHDAT-Car corpus and an improvement with the Loquendo recognizer on SNOW-Factory corpus.