Evaluating models of vowel perception.

Abstract

There is a long-standing debate concerning the efficacy of formant-based versus whole spectrum models of vowel perception. Categorization data for a set of synthetic steady-state vowels were used to evaluate both types of models. The models tested included various combinations of formant frequencies and amplitudes, principal components derived from excitation patterns, and perceptually scaled LPC cepstral coefficients. The stimuli were 54 five-formant synthesized vowels that had a common F1 frequency and varied orthogonally in F2 and F3 frequency. Twelve speakers of American English categorized the stimuli as the vowels /I/, /[symbol: see text]/, or /[symbol: see text]/. Results indicate that formant frequencies provided the best account of the data only if nonlinear terms, in the form of squares and cross products of the formant values, were also included in the analysis. The excitation pattern principal components also produced reasonably accurate fits to the data. Although a wish to use the lowest-dimensional representation would dictate that formant frequencies are the most appropriate vowel description, the relative success of richer, more flexible, and more neurophysiologically plausible whole spectrum representations suggests that they may be preferred for understanding human vowel perception.