Modeling Individual Differences in Categorical Perception with a Bayesian Framework

Listeners vary in their perception of speech, falling along a continuum from categorical to continuous. We applied a Bayesian computational framework to model this individual difference in speech perception. We analyzed publicly available data (Honda et al., 2024) from 195 participants across four phonetic conditions using both two-alternative forced choice and visual analogue scale tasks. Our model characterizes each listener’s perception using two key parameters: perceptual warping (τ), the signal-to-noise ratio of phonetic encoding, and noise variance (\( \sigma_S^2 \)), a proxy to perceptual noise in experimental designs. Combining these two parameters revealed four perceptual profiles: Veridical (high τ, low \( \sigma_S^2 \)), Categorical (low τ, low \( \sigma_S^2 \)), Compensatory (low τ, high \( \sigma_S^2 \)), and Noisy (high τ, high \( \sigma_S^2 \)). These profiles predicted behavioral patterns coherently, while successfully distinguishing between listeners who would appear similar when characterized by behavioral measures alone. Critically, results revealed that profile distributions shifted dramatically based on phonetic conditions, with primary cues yielding a balanced mix of profiles and secondary cues producing distributions skewed heavily toward Veridical and Compensatory listeners (80%). Underscoring this flexibility, intraclass correlations for both τ and \( \sigma_S^2 \) ​ were zero, with phonetic condition effects 30 times stronger for \( \sigma_S^2 \) (χ² = 803.91) than τ (χ² = 29.47). These findings challenge the traditional view of categorical perception as a fixed characteristic, demonstrating instead that it is a flexible, context-driven perceptual state.