The TRACE Model of Speech PerceptionPresented by A. Brian Davis• Introduction• Specifics• Experiments• LimitationsIntroduction●Insufficiency acoustic information−Context Sensitivity (adjacent phonemes)●Distributed processing−Individual units as hypotheses●Psychological vs Computational model performance• Introduction• Specifics• Experiments• LimitationsSpecifics●Input Features−Consonantal, vocalic (obvious), diffuseness (spread out), acuteness, voicing, Power, Amplitude of “burst of noise” (discriminating stops)−Discretized, chosenSpecifics●Input Features−Consonantal, vocalic (obvious), diffuseness (spread out), acuteness, voicing, Power, Amplitude of “burst of noise” (discriminating stops)−Discretized, chosen●CorrelatedTRACE 2●Time discretized●3 logical levels−Features, Phonemes, Words●Time (1,3,6); overlap●Node excitation is hypothesis−Probabilistic: , −{Between,Within}-level excitation, inhibition (Neural network)●Features excite phoneme nodes●Phonemes excite word nodes●Competitionp Ri=Si∑jSjSi=ekai• Introduction• Specifics• Experiments• LimitationsjTRACE●Java implementation TRACE●Thanks to Daniel Felps for the linkjTRACE●Java implementation TRACE●Thanks to Daniel Felps for the link●Somewhat faulty−ExperimentsInput phonemesLexical Effects●Correct misspoken phoneme−Word feedback (Ex: Trust/Tlust)●Mechanical / Psychological●Reaction times (Phonotactic effect)−End of words (secret, guldut)Word Segmentation●Ambiguous word breaks−Humans easily disambiguate (usually)●secant/she can't−TRACE words inhibit other words overlapping time●Shorter words preferred, then longer−Words vs non-words, break decision speed●{Possible, pagusle} target• Introduction• Specifics• Experiments• LimitationsDeficiencies●Context●Size−Learning mechanism
View Full Document