Measuring articulatory similarity with algorithmically reweighted principal component analysis.

Articulatory similarity was assessed for a corpus of 2700 ultrasound images of 75 cross‐linguistically frequent speech sounds produced by four subjects in three segmental contexts (a_a, i_i, u_u). Cross‐distances were generated for the entire length of the vocal tract using the Palatron algorithm and realistic estimates of the location of the pharyngeal wall and teeth, resulting, after interpolation, in 60 cross‐distances per token. A standard principal component analysis of these data is overwhelmed by the coarticulatory effects of the context vowels. Algorithmically reweighted principal component analysis was devised in order to use coarticulatory variation to isolate the most distinctive cross‐distances for each target segment. The reweighting algorithm considers the variance across repetitions in each of the 60 cross‐distances, as well as variance in the slope of the cross‐distance function, in order to identify areas of stability across tokens. For each target sound, cross‐distances with the greatest...