关键词: 语音感知, 说话人归一化, 元音, 塞音, 线性弹道累加器模型

Abstract: Speech perception is influenced by contextual factors, which significantly shape how listeners map acoustic cues to linguistic categories. This phenomenon is often revealed through speech categorization tasks, yet the precise cognitive-computational processes underlying these contextual effects remain underexplored. Previous research has established that speaker-specific characteristics, such as gender, systematically modulate speech perception, well known as talker normalization. However, a comprehensive understanding of how talker normalization is integrated into the real-time perceptual decision-making process is still lacking. This study applied a linear ballistic accumulator (LBA) model to analyze listeners’ behavioral data from vowel and stop continuum tasks, aiming to elucidate the computational cognitive mechanisms of talker normalization.
To this end, we conducted four experiments, testing how contextual gender influences the perception of vowel formants and stop voice onset time (VOT). Experiments 1a and 1b used male and female nine-step vowel continua (ranging from [a] to [e]), respectively, while Experiments 2a and 2b employed male and female nine-step stop continua ([k] - [k^h]). In each trial, a continuum stimulus was preceded by a sentence spoken by either a male or a female talker, thereby manipulating the contextual gender. Listeners provided a two-alternative forced-choice judgment, and we collected their categorical responses and reaction times for analysis.
Our analysis, utilizing an LBA model that allowed contextual gender to modulate decision threshold (b), non-decision time (t₀), and drift rate (v), provided the best fit to the data across all four experiments. The modulation of drift rate by contextual gender was consistent with the pattern observed in participants' categorical responses, revealing assimilation effects for vowel perception and compensation effects for stop perception. Furthermore, female contexts consistently prolonged non-decision time and reduced decision threshold compared to male contexts in Experiments 1a, 2a, and 2b. However, these effects were absent for the female vowel target (Experiment 1b), suggesting that target talker gender modulates the expression of contextual effects in a feature-specific manner.
These results provide computational evidence for multi-stage cognitive mechanisms underlying contextual effects on speech perception, as revealed by the LBA framework. By integrating categorical responses and reaction times, the LBA model allowed us to pinpoint specific subprocesses affected by context—sensory encoding (t₀), evidence accumulation efficiency (v), and decision criterion (b)—going beyond simply observing shifts in categorization functions. The differential modulation of these latent processes by contextual gender for formants and VOT suggests a feature-specific, multifaceted mechanism by which environmental factors influence speech perception. Crucially, this evidence-accumulation framework provides a powerful tool for disentangling the complex interplay between acoustic cues and cognitive processes, enhancing our understanding of how contextual variability shapes flexible speech categorization across diverse listening environments. This constitutes a finer-grained account of contextual effects than traditional analyses based solely on categorization shifts.

Key words: speech perception, talker normalization, vowel, stop, linear ballistic accumulator model