We defined COSMO, a “family” of Bayesian models simulating some of the fundamental properties of speech perception and production. COSMO makes it possible to study situations where the auditory and motor systems are redundant (indistinguishability theorem) or, on the contrary, complementary (“auditory-narrow, motor-wide” property). COSMO accounts for the learning of representations, the resulting idiosyncrasies, and their perceptual-motor couplings. The variant of COSMO which deals with speech production (Bayesian-GEPPETO) also accounts for adaptation patterns in response to disturbances. Our goal is to bring COSMO to process real continuous speech, and to develop a hierarchical processing architecture faithful to neural representations.
Marc-Antoine Georges,Mamady Nabe