Factor analysis of vocal-tract outlines derived from real-time magnetic resonance imaging data

A factor analysis of vocal-tract outlines derived automatically from real-time magnetic resonance image (rtMRI) sequences has been performed. The analysis results in a compact representation of vocaltract shapes, where every utterance is represented by a small set of trajectories corresponding to weights in linear combinations of linguistically interpretable vocal-tract deformations. Vocal-tract shapes can be reconstructed with good accuracy from these trajectories. The work uses information from a significantly larger number of speech frames compared to previous attempts in articulatory modeling. The proposed method is illustrated through a case study of rtMRI data corresponding to 250 sentences spoken by a single speaker and underscores the promise of the methodology for phonological analysis and articulatory synthesis.