A comparative study of spectral peaks versus global spectral shape as invariant acoustic cues for vowels

The primary objective of this study was to compare two sets of vowel spectral features, formants and global spectral shape parameters, as invariant acoustic cues to vowel identity. Both automatic vowel recognition experiments and perceptual experiments were performed to evaluate these two feature sets. First, these features were compared using the static spectrum sampled in the middle of each steady-state vowel versus features based on dynamic spectra. Second, the role of dynamic and contextual information was investigated in terms of improvements in automatic vowel classification rates. Third, several speaker normalizing methods were examined for each of the feature sets. Finally, perceptual experiments were performed to determine whether vowel perception is more correlated with formants or global spectral shape. Results of the automatic vowel classification experiments indicate that global spectral shape features contain more information than do formants. For both feature sets, dynamic features are superior to static features. Spectral features spanning a time interval beginning with the start of the on-glide region of the acoustic vowel segment and ending at the end of the off-glide region of the acoustic vowel segment are required for maximum vowel recognition accuracy. Speaker normalization of both static and dynamic features can also be used to improve the automatic vowel recognition accuracy. Results of the perceptual experiments with synthesized vowel segments indicate that if formants are kept fixed, global spectral shape can, at least for some conditions, be modified such that the synthetic speech token will be perceived according to spectral shape cues rather than formant cues. This result implies that overall spectral shape may be more important perceptually than the spectral prominences represented by the formants. The results of this research contribute to a fundamental understanding of the information-encoding process in speech. The signal processing techniques used and the acoustic features found in this study can also be used to improve the preprocessing of acoustic signals in the front-end of automatic speech recognition systems.