Formant-based articulatory strategies: Characterisation and inter-speaker variability analysis

Abstract

Vowels are articulatorily characterised by the shape of the vocal tract and acoustically by their three lowest formants. The relationship between formant variations and articulatory variations is well documented. This study addresses the opposite problem: describing the main articulatory variations associated with the variations of single formants. A data-driven modelling-based approach was chosen for this purpose. Midsagittal vocal tract contours from the glottis to the lips for 532 vowels from 41 speakers of three different languages were obtained from MRI data. Corresponding formant values were obtained by acoustic modelling. For each speaker, linear regressions of the contours on the formant values were performed. It led to five articulatory components, characterising the vocal tract variations associated with variations of the first three formants and their differences. Inter-speaker variability was analysed by applying principal components analysis on the components in a second level of modelling. A correlation analysis of the resulting inter-speaker components with morphological features was performed to determine whether a speaker’s strategy could be driven by the morphology. Results show that the palate shape and the vertical pharyngeal height, related to the male–female difference, have a small influence on the speaker’s strategy. Associated Matlab code is publicly available.