Pistes pour le contrôle d'un robot parlant capable de réduction vocalique

Abstract

The purpose of this thesis is to study how the phonemic command, abstract and invariant in nature, is coded in terms of motor control, to generate a physical and variable acoustic signal. A motor control model is exploited, based on the Equilibrium Point hypothesis, proposed by A. Feldman, which suggests that movement results from shifts in variables controlling the system’s mechanical equilibrium. The hypothesis studied here consists in relating discrete motor commands, specifying target equilibrium points, to phonemes. Realistic articulatory and acoustic trajectories were generated with this model for Vowel-Vowel-Vowel sequences, uttered by a French speaker, under several speech conditions. Articulatory trajectories were inferred from the acoustic signal by a first inversion involving an articulatory model of the vocal tract. A functional dynamic model of speech articulators, applied to the tongue body articulator, which is considered as the most relevant in the studied sequences, is then inverted to recover the motor commands underlying the trajectory observed in the “ideal” speech condition. Then from the motor commands thus inferred for the ideal condition, articulatory and acoustic signals, similar to the ones observed in other conditions, are generated by simply tuning appropriate dynamic parameters. A new track is thus open for adaptive speech synthesis. Sensitivity analysis and perceptual tests on synthetic signals attest for the model’s ability to simulate focus stress and rate effects using constant postural (equilibrium point) commands.