Can static vocal tract positions represent articulatory targets in continuous speech? Matching static MRI captures against real-time MRI for the French language

This paper uses mediosagittal slices of a static magnetic resonance imaging (MRI) dataset capturing the blocked articulation of vowels and of consonants that anticipate /a, i, u, y/ and a variety of other vowels to study the presence and distinctness of these deliberately taken articu-latory targets in real-time MRI recordings. The study investigates whether such articulatory targets are actually attained in fluent speech, how marked they are, and what factors influence the degree of similarity between a given articulatory target and the actual vocal tract shape. To quantify the similarity, we use structural similarity, Wasserstein distance, and SIFT measure. We analyze the amplitude and timing of the observed similarity peaks across different phonetic classes and speech types (spon-taneous versus not). We show that although real-time speech involves shapes quite similar to the static data, there is a great intra-and inter-speaker variability.