The FACTS model: using state estimation and task-based feedback control to model the speech motor system

Wepresentanewcomputationalmodelofspeechmotorcontrol:theFACTSmodel(Feedback-Aware ControlofTasksinSpeech).Thismodelemploysahierarchicalarchitecture,withcontrolofhigh-levelspeechtasksseparateandabovethecontrolofspeecharticulatorpositions.Thetask-levelcontrollerismodeledasadynamicalsystemgoverningthecreatingofdesiredconstrictionsinthevocaltract,drawingfromtheTaskDynamicsmodel.Critically,thistask-levelfeedbackcontrollerusesanestimateofthecurrentstateofthevocaltractratherthanthetruestatetogeneratemotorcommands.Thisstateestimateisbasedonaninternalpredictionofthestateaswellasauditoryandsomatosensoryfeedback.WeshowthattheFACTSmodelisabletoqualitativelyreplicatemanycharacteristicsofthehumanspeechsystem:themodelisrobusttonoiseinboththesensoryandmotorpathways,isrelativelyunaffectedbyalossofauditoryfeedbackbutimpactedtoalargerdegreebythelossofsomatosensoryfeedback,andrespondsappropriatelytoexternally-imposedalterationsofauditoryandsomatosensory

feedback.Themodelalsoreplicatespreviouslyhypothesizedtrade-offsbetweenrelianceonauditoryandsomatosensoryfeedbackinspeechmotorcontrolandshowsforthefirsttimehowthisrelationshipmaybemediatedbyacuityineachsensorydomain.Theseresultshaveimportantimplicationsforour

understandingofthespeechmotorcontrolsysteminhumans.