Analyzing speech in both time and space: Generalized additive mixed models can uncover systematic patterns of variation in vocal tract shape in real-time MRI

• DOI: 10.5334/labphon.214
• Publication Date: 2020-03-18T12:42:47Z
• Authors: Christopher Carignan; Phil Hoole; Esther Kunay; Marianne Pouplier; Arun Joseph; Dirk Voit; Jens Frahm; Jonathan Harrington
• Keywords: Language. Linguistic theory. Comparative grammar; P101-410; real-time MRI; GAMM; FLMM; speech dynamics; speech imaging; speech dynamics; speech imaging; speech processing systems; GAMM; 03 medical and health sciences; 0302 clinical medicine; XXXXXX - Unknown; oral communication; magnetic resonance imaging; real-time MRI; FLMM; Articulatory Phonetics

We present a method of using generalized additive mixed models (GAMMs) to analyze midsagittal vocal tract data obtained from real-time magnetic resonance imaging (rt-MRI) video of speech production. Applied to rt-MRI data, GAMMs allow for observation of factor effects on vocal tract shape throughout two key dimensions: time (vocal tract change over the temporal course of a speech segment) and space (location of change within the vocal tract). Examples of this method are provided for rt-MRI data collected at a temporal resolution of 20 ms and a spatial resolution of 1.41 mm, for 36 native speakers of German. The rt-MRI data were quantified as 28-point semi-polar-grid aperture functions. Three test cases are provided as a way of observing vocal tract differences between: (1) /aː/ and /iː/, (2) /aː/ and /aɪ/, and (3) accentuated and unstressed /aː/. The results for each GAMM are independently validated using functional linear mixed models (FLMMs) constructed from data obtained at 20% and 80% of the vowel interval. In each case, the two methods yield similar results. In light of the method similarities, we propose that GAMMs are a robust, powerful, and interpretable method of simultaneously analyzing both temporal and spatial effects in rt-MRI video of speech.