Neural Control Of Physiological Articulatory Model

Existing neurocomputational approaches use geometrical vocal tract models tosimulate processes of speech production. These approaches currently are not capableof replicating the biomechanical process of speech production in detail. In this study,we implemented a physiological articulatory model which is designed to faithfullyrealize the biomechanical mechanism of speech production into a neural controlenvironment. Neural representations for high-level motor, low-level motor,somatosensory, as well as auditory states of speech items were defined as the interfacebetween physiological articulatory model and neurocomputational controlenvironment. Babbling training was carried out to validate our neural control model.Simulation results showed that our model is able to realize vocalic dimensions andbuild up the association between motor and sensory states.We proposed two modeling hypotheses for our neural model, both of ourhypothetic neural control models include feedforward and feedback sensorimotorpathways and a self-organizing map (SOM) bridged between the pathways. Twosimulation experiments have been conducted for implementing pre-linguistic speechknowledge within the two versions of the model. Thus we can conclude from theresult that there is a need for strict separation of high-level and low-level motor andsensory state representations and it can be concluded from this study as well thatdifferent neural processing modules should be postulated for these two levels ofcontrol. Self-organizing map related to the contour positions of control points andmuscle activation patterns was established during speech motor learning. As SOMlearning leads to meaningful results only in the case of motor extention modelhypothesis, experimental result refer to the one-to-many problem in the mappingbetween the high-level to the low-level motor states, which indicates that quitedifferent muscle activation patterns can lead to similar articulatory positioning.