A mixed generative model of auditory word repetition

Abstract

In this paper, we introduce a word repetition generative model (WORM), which--when combined with an appropriate belief updating scheme--is capable of inferring the word that should be spoken when presented with an auditory cue. Our generative model takes a deep temporal form, combining both discrete and continuous states. This allows a (synthetic) WORM agent to perform categorical inference on continuous acoustic signals, and--based on the same model--to repeat heard words at the appropriate time. From the perspective of word production, the model simulates how high-level beliefs about discrete lexical, prosodic and context attributes give rise to continuous acoustic signals at the sensory level. From the perspective of word recognition, it simulates how continuous acoustic signals are recognised as words and, how (and when) they should be repeated. We establish the face validity of our generative model by simulating a word repetition paradigm in which a synthetic agent or a human subject hears a target word and subsequently reproduces that word. The repeated word should be the target word but differs acoustically. The results of these simulations reveal how the generative model correctly infers what must be repeated, to the extent it can successfully interact with a human subject. This provides a formal process theory of auditory perception and production that can be deployed in health and disease. We conclude with a discussion of how the generative model could be scaled-up to include a larger phonetic and phonotactic repertoire, complex higher-level attributes (e.g., semantic, concepts, etc.), and produce more elaborate exchanges.