Investigating patterns of vowel production using real-time magnetic resonance imaging

Characteristic tongue postures of 21,478 vowels produced by four speakers of General American English were compared in lexically stressed and unstressed positions to examine patterns of articulation and reduction. Image frames capturing midsagittal vocal tract configurations were reconstructed at the acoustic midpoint of each vowel produced in a 3,450 word forced-aligned transcribed speech corpus (Narayanan et al. 2014). Sixty image sets were constructed, consisting of all exemplars of each contrastive vowel quality of American English, as produced by each speaker: up to 1049 tokens of the most frequent vowel /i/, and 48 tokens of the least frequent vowel /ɔɪ/. Image frames were cropped to a region of interest centred on the tongue dorsum, and linearized into one-dimensional vectors in which pixel intensity corresponds to presence of soft tissue. Principle components analysis was applied to each set of image vectors to examine the distribution and variability of tongue tissue associated with each vowel posture. “Eigentongues” were constructed from the most significant components in each image set (Sirovich & Kirby 1987; Hueber et al. 2007), resulting in a set of images characterizing each speakers’ vowel space, produced in both lexically stressed and unstressed positions.