Phoneme Perception as Bayesian Inference with a Narrow-Tuned Multimodal Prior

Abstract

Statistical inference well describes sensory perception: estimating true status of the world according to newly obtained sensory information and beliefs that are formed by previous experience. Beyond previous attempts of explaining perception by Bayesian framework, distribution of prior is inferred by fitting a Bayesian model to perceptual bias and variability exhibited by observers (Girshick et al., 2011). In current study, we inferred a prior that is being combined with given sensory information in phoneme perception, where the presence of a strong prior is expected. For inferring the prior, subject performed two distinctive psychophysics experiments: identification and discrimination. The acoustic stimuli varied gradually along the spectrum encompassing three different stop consonant - /ba/, /da/, and /ga/. A significant component of model, which is prior, is estimated as mixture of three normal distribution having the means and variance of which reflect the centers and spread of phoneme stimuli that is most frequently heard by the listener in the past. Likelihood is similarly modeled as normal distribution except having its mean corresponding to given stimuli and variance identical to all types of stimuli. Only with a few numbers of free parameters, the hallmark features of phoneme perception are well explained simultaneously: drastic change of selection category in identification task and enhanced discriminability around boundaries of two phonemes. Further in goodness of fit, our model implementing mixture normal surpassed a model with uniform prior distribution and matched with a model having non-parametric prior. Suggested Bayesian model provides evidence that human phoneme perception requires a narrow-tuned multimodal prior whose peak exists at prototypical phoneme stimuli.