Tracking sources of formant frequency variation in data from a time-controlled speech production task

Vowel formant structure is particularly sensitive to influence from neighboring segments in connected speech. Statistical models can successfully account for acoustic variability from multiple simultaneous sources of variance, such as anticipatory effects of following consonants and vowels (see, for example, Cole et al., 2010). Statistical approaches have also been developed to predict context segments from static measures of spectral variation like mean or midpoint formant frequencies. This study examines shifts in the relative strength of these sources of variance over the course of the target vowel, and compares previous approaches to predicting the identity of upcoming segments with a novel approach which is cumulatively informed by formant frequency variation as it unfolds temporally. Data are from 32 participants who produced target-context word sequences (permutations of …V1C#hV2…, where V1 = [ɛ], C = [θ f], and V2 = [i ɛ æ ɑ u], e.g., “deaf-heating”) within a carrier sentence. Timing and prosody were strictly controlled by entraining the syllable rate to a metronome. Context word onset was always [h], which has no intrinsic supralaryngeal articulation and only begins foot-initial syllables in English. This ensured a consistent prosodic boundary between target and context words and prevented resyllabification of C across words.