Predictive Neural Computations Support Spoken Word Recognition: Evidence from MEG and Competitor Priming

Area of Science:

• Neuroscience
• Psycholinguistics
• Computational Auditory Neuroscience

Background:

• Human speech comprehension relies on Bayesian inference, combining sensory input with prior knowledge.
• The neural mechanisms of Bayesian perceptual inference in spoken word recognition are not fully understood.
• Existing models propose competitive selection (e.g., TRACE) or predictive selection (e.g., Predictive-Coding).

Purpose of the Study:

• To investigate the neural implementation of Bayesian perceptual inference in spoken word recognition.
• To differentiate between competitive-selection and predictive-selection accounts using magnetoencephalography (MEG).
• To examine the role of prediction error in updating lexical representations during word identification.

Main Methods:

• Collected MEG data from human listeners (male and female).
• Employed a competitor priming manipulation to alter word prior probabilities.
• Analyzed neural responses in relation to lexical decision times and points of unique phonetic identification.

Main Results:

• Lexical decisions showed delayed recognition of target words following a related prime word.
• MEG responses in the superior temporal gyrus (STG) showed increased activity after unique phonetic identification for word primes.
• These effects were absent for pseudoword primes and targets, indicating sensitivity to lexical status.

Conclusions:

• Findings support predictive-selection theories, where prediction error updates lexical probabilities.
• Neural computations of prediction error play a crucial role in spoken word recognition.
• This contrasts with competitive-selection models that rely solely on direct inhibition between word representations.