Evaluating models of vowel perception.

Area of Science:

• Auditory Perception
• Acoustic Phonetics
• Speech Science

Background:

• The debate on vowel perception models centers on whether formant frequencies or whole spectrum analysis is more effective.
• Previous models often relied on simplified acoustic features, potentially missing crucial perceptual information.

Purpose of the Study:

• To evaluate the efficacy of formant-based versus whole spectrum models in predicting human vowel categorization.
• To determine the best acoustic representation for understanding steady-state vowel perception.

Main Methods:

• Synthetic steady-state vowels with controlled formant frequencies (F1, F2, F3) were created.
• American English speakers categorized these vowels.
• Categorization data were used to test models based on formant frequencies, principal components of excitation patterns, and LPC cepstral coefficients.

Main Results:

• Formant frequencies alone were inadequate; nonlinear terms (squares, cross-products) improved their predictive power.
• Principal components derived from excitation patterns provided reasonably accurate fits.
• Whole spectrum representations showed greater success than simple formant frequency models.

Conclusions:

• While formant frequencies offer a low-dimensional description, they require nonlinear extensions for accurate vowel perception modeling.
• Richer, whole spectrum representations are more flexible and neurophysiologically plausible.
• Whole spectrum models may be preferred for a more complete understanding of human vowel perception.