Predicting Choices and Timings of Phoneme Categorization with a Perceptual Decision Model of Phonemic Processing

Abstract

Despite crucial roles of pre-lexical units in speech perception, modeling efforts so far have been heavily focused on information processing at lexical or post-lexical stages, impeding the mechanistic investigation of speech perception. Given this dearth of frameworks for studying pre-lexical units, the current study proposes a system-level neural model for phoneme classification. A lynchpin idea behind the proposed model is that the brain represents phonemes as probabilistic quantities, likelihoods. With this idea, our model bridges three well-known canonical computations in the brain – sensory encoding, likelihood decoding and evidence accumulation - along a cascade hierarchy of neural processing towards generating inputs to a next stage of speech perception. At the initial stage, sensory neurons with different tuning curves for physical properties relevant to phoneme discrimination compute individual likelihoods for the presence of those properties. Phoneme neurons at the following stage compute likelihoods for specific phonemes by summing the outputs of those sensory encoding neurons with weighting curves tuned for their preferred phonemes. At the final stage, evidence-accumulation neurons compute and accumulate over time evidence to reach a discrete phoneme classification by integrating outputs of phoneme neurons in a task-optimal manner over time. The accumulation-to-bound mechanism operating at this stage translates probabilistic information represented in the phoneme neurons output into concrete choices at a certain time. This translation allowed us to test the empirical viability of our model by assessing its capability of predicting actual patterns of choice fractions and reaction times exhibited by human listeners engaging in phoneme classification under various listening conditions. Using a small number of parameters, the model predicted not only the static, categorical structure of phoneme classification as a function of physical stimulus property, but also the adaptation-induced, dynamic changes in classification on an identical stimulus. Furthermore, the model was flexible enough to cover the wide range of individual differences in phoneme classification behavior. With these behavioral constraints in conjunction with neural and computational constraints exercised in model construction, our model provides a framework for studying neural mechanisms underlying initial stages of speech processing by generating hypotheses and predictions that are testable by neurophysiological and behavioral experiments.