Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Sentence Predictability Shapes the Encoding of Phonetic Detail
Poster Session A - Sandbox Series, Thursday, October 24, 10:00 - 11:30 am, Great Hall 3 and 4
This poster is part of the Sandbox Series.
Will Chih-Chao Chang1, Jiaxuan Li1, Xin Xie1; 1University of California, Irvine
It is well established that listeners utilize subtle phonetic detail (e.g., voice onset time, VOT) to distinguish between words in isolation (e.g., GOLD vs. COLD; McMurray et al., 2002). However, when given a constraining sentence context, how top-down predictions from prior contexts affect the encoding of phonetic detail remains unclear. Some studies found listeners encode phonetic detail of words less accurately in high-predictability sentence contexts compared to low-predictability ones (Manker, 2019), while others found high-predictability sentence contexts enhance acoustic-phonetic processing (Broderick et al., 2019). Furthermore, most studies presented target words in isolation or at sentence end, leaving open how phonetic encoding and prior prediction interact during ongoing speech. Our study investigates whether listeners dynamically adjust the precision with which they encode phonetic detail as speech unfolds. Specifically, we examine how sentence predictability affects the phonetic encoding of English word-initial voiced stops (e.g., /g/ in GOLD). VOT is manipulated to create auditory voiced targets with three levels of phonetic ambiguity, which would gradiently activate voiced targets and their voiceless competitors (e.g., GOLD 100%-COLD 0%, GOLD 75%-COLD 25%, GOLD-50%-COLD 50%). The auditory voiced targets appear in the middle of sentences with high-predictability (e.g., The treasure hunter found a gold necklace) or low-predictability (e.g., The young man found a gold necklace) contexts. To obtain a fine-grained predictability measure, we further calculated the word surprisal (negative log probability) of the voiced targets given the preceding sentence context using large language model GPT2 (Radford et al., 2019). Participants listen to auditory sentences containing the voiced targets while performing a lexical decision task, where visual strings appear at the offset of target words. There are 72 critical trials where three types of visual strings are presented: identical voiced targets (e.g., GOLD), their voiceless competitors (e.g., COLD), or controls (e.g, ANT), along with 144 filler trials paired with controls or non-words. Participants continue listening to the sentences while responding. Probe trials, where listeners answer comprehension questions, will be randomly interleaved to ensure attention to the auditory sentences. The encoding precision of phonetic detail in the VOT of auditory voiced targets is hypothesized to affect the reaction times (RTs) in the lexical decision task. GLMM models will be fit to the RTs data, with phonetic ambiguity, sentence predictability, visual string type and their interactions as the fixed effects. We predict more accurate encoding of phonetic detail in unpredictable context, as evidenced by a three-way interaction: increased phonetic ambiguity of VOT slows RTs for visual voiced targets but speeds RTs for voiceless competitors, with a stronger effect in low- than high-predictability sentences. We also expect a stronger VOT effect on RTs as word surprisal increases, suggesting phonetic encoding is sensitive to fine-grained changes in sentence predictability. In the future, we will use fMRI to investigate neural correlates underlying the encoding of phonetic detail in STG, and how STG activation interacts with the cortical regions sensitive to word surprisal. These findings will further elucidate the dynamics of bottom-up and top-down processing within the bilateral fronto-temporo-parietal network during speech comprehension.
Topic Areas: Speech Perception,