Event Date
Speaker: Santiago Barreda Castañón, Associate Professor, Linguistics, UC Davis
Title: "Investigating formant repetition error using the Davis Repetition Corpus"
Abstract: Formants are the peaks in the spectral envelope of vowel sounds, and most strongly determine their vowel quality. Statistical analyses of vowel formants (e.g., ANOVA, linear regression) typically assume constant (homoscedastic) error variance across all vowels and formants. However, this assumption is rarely tested empirically. In this study, we analyze within-speaker formant variability using a large corpus of repeated vowel productions to determine the structure of production error. Results suggest repetition error is neither constant in Hertz (as assumed by standard models) nor directly proportional to frequency (as assumed by log-transformed models). Instead, error standard deviation scales with mean frequency following a power law with an exponent of approximately 0.54. This scaling of repetition errors is inconsistent with constant articulatory precision. Rather, it suggests that articulation of high-frequency formants is disproportionately accurate, a result that may indicate the reduction of production error using tactile feedback between articulators for smaller resonance cavities. Additionally, we find weak correlations between errors across formants for most vowels, the notable exception being a positive correlation between F2 and F3 for front vowels. As a result, the assumption of homoscedasticity is empirically unfounded and articulatorily implausible, and it is recommended that phonetic analyses employ heteroscedastic regression models.
Faculty website: (https://linguistics.ucdavis.edu/people/santiago-barreda)