Self-perception and vowel inherent spectral change

Speakers self-correct vowels with deviant formants: tokens whose onsets are farther from the average midpoint move toward that average, while nearer tokens move randomly (Niziolek et al., 2013). The method used to establish this finding typically groups tokens based on onset formants relative to midpoint formants, and the midpoint is taken as the target. But vowel inherent spectral change research has shown that vowels are best modeled as contours rather than midpoints. The present study asks whether speakers self-correct toward trajectories rather than static midpoints. Sixty-eight speakers provided 20 tokens of each stressed non-rhotic monophthong in English in [hVd] (heed, hid, etc.). Formants were sampled at 3% and 50% of vowel duration, and each speaker’s average formant position was calculated at both samples. Sample-matched Euclidean distances were calculated — each 3% token to the 3% average, each 50% token to the 50% average — rather than comparing to the midpoint sample in both cases. Tokens were grouped based on onset distance. Far tokens showed a larger decrease in distance by the midpoint than Near tokens, and regression to the mean was ruled out as the reason. This study suggests that speakers use dynamic spectral information for online self-correction.