Perceptual speech processing and phonetic feature mapping for robust vowel recognition

Abstract

We propose perceptual speech processing and phonetic feature mapping, which are inspired by the human auditory perceptual characteristics. The proposed perceptual speech processing is based on three perceptual characteristics and consists of three independent processing steps: masking effect, minimum audible field renormalization, and mel-scale resampling. They remove unperceptible spectral components, and adjust the magnitude and frequency scales of speech spectra, respectively. We apply these three processing steps to the speech spectrum sequentially to generate a new speech signal representation called the perceptual spectrum. For Mandarin vowel recognition, nine representative vowels are selected as references and similarity measures to these reference spectra, called phonetic features, are then generated from the perceptual spectrum. These phonetic features then serve as speech parameters in a continuous HMM-based recognition, stage. With these two techniques, a high recognition accuracy on Mandarin vowel phonemes has been achieved. Further experiments confirm that significant improvement on recognition robustness with respect to speaker variation and noise contamination can also obtained.