Speaker: Gasser Elbanna (Harvard)
Title: A model of speech recognition reproduces behavioral signatures of human speech perception and reveals mechanisms
Time: Monday, November 3rd, 5pm – 6:30pm
Location: 32-D831
Abstract: Humans dexterously extract meaning from variable acoustic signals and can faithfully repeat back novel utterances—hallmarks of spoken communication. Speech perception is thought to subserve these downstream tasks via transforming sound into robust perceptual representations. Yet progress on the nature of these representations and their mechanisms has been limited by the lack of (i) stimulus-computable models that replicate human behavior and (ii) large-scale behavioral benchmarks for comparing model and human speech perception. In this talk, I will present our work on developing candidate artificial neural network models of human speech perception along with new behavioral experiments to compare phonetic judgments in humans and models. Our models reproduce patterns of human responses and confusions alongside recapitulating key behavioral signatures of human speech perception. I will also show how our models enable us to investigate the role of contextual integration and its directionality in speech perception.