ACOUSTIC CORRELATES OF PHONETIC PERCEPTION

点击翻译按钮获取中文摘要

Numerous practical and theoretical problems could be addressed if we had a better understanding of the auditory mechanisms underlying phonetic recognition. This proposal is aimed at improving our understanding of these mechanisms, with a particular focus on vowel perception. Although there is a long tradition of representing vowels by the spectral pattern sampled at a single time slice, a growing body of literature suggests that dynamic properties play an important role in vowel identification. Despite this literature, relatively little is known about the precise mechanisms that are involved in mapping dynamic spectral cues onto perceived vowel quality. Some of the proposed experiments will test specific hypotheses about the way in which this mapping might occur. The experiments will make use of a large database consisting of vowels spoken by 150 talkers (men, women, and children). Measurements of fundamental frequency (F0) and formant contours from these signals will be used in a series of experiments designed to determine the role played by F0, vowel duration, and spectral change in vowel identification. Specific hypotheses will be tested by: (1) acoustic analysis of tokens in the 150-talker database, and (2) listening tests involving various kinds of stimuli resynthesized from these tokens. A second goal of this project is to evaluate the "Masked Peak Representation" (MPR), a new method of representing speech which was developed as an alternative to both traditional formant representations and "whole spectrum" representations. Formant representations are widely used because these they can account for a relatively large number of findings in phonetic perception. The principal weakness of formant theory is that tracking formants in natural speech is a difficult and essentially unresolved problem. Largely in response to this problem, some investigators have proposed a whole spectrum approach in which phonetic quality is controlled by overall spectral shape. The whole spectrum approach, however, cannot account for very convincing data showing that judgments of phonetic quality are affected primarily by the frequencies of spectral peaks, and relatively unaffected by spectral shape details in nonpeak regions. The MPR was designed to retain maximal sensitivity to spectral peaks but without requiring the explicit tracking of formants. The basic idea behind the MPR is to: (1) obtain a pitch-independent spectrum through cepstral smoothing, (2) stimulate nonlinear auditory frequency coding by computing a bark-scale transform, and (3) simulate lateral suppression by subtracting a running average of spectral values. The resulting "masked spectrum" retains spectral peaks but removes most other spectral shape details. The MPR will be evaluated with: (1) an experiment comparing MPR-based predictions of perceived phonetic distance with those of a more traditional auditory model, (2) speech recognition tests that use a Hidden Markov Model to map sequences of MPR spectra onto either words of phonetic segments, and (3) listening tests with speech resynthesized from MPR spectra.