Towards large-scale cross-speaker articulatory modeling of vowels

While previous studies have attempted to decompose vowel articulatory data into a set of basis factors, these studies have often been limited in both scale and the data being sparsely sampled, limiting interpretability and generalizability of the results (Nix et al. 1996 and Serrurier et al. 2019). In this study, the data were analyzed from 36 (23F, 13M) American English speakers producing 13 vowels in bVt sequences obtained using real-time MRI. Midsagittal tongue contours were obtained during vowel productions for all speakers using a semi-automated segmentation algorithm (Jain et al. 2024). Frames corresponding to the vowel articulation were segmented using MFA and simultaneously recorded audio. A combination of Procrustes analysis for cross-speaker normalization and guided PCA were employed to decompose the pooled articulatory space into a set of vowel “primitives.” 71% of the variation within the articulatory space is captured in the first three principle components, with the first component representing variation in tongue body advancement/retraction, the second component fronting/backing, and the third lesser degrees of fronting/backing. This pattern is consistent when the data is split into male/female groups. Vowel-wise PCA loadings for the first two components mostly show expected degrees of separation between the dimensions of advancement/retraction and backness, though some distinctions appear to be diluted, likely attributable to the normalization procedure. [Work supported by the NIH.]