A core challenge for theories of human speech recognition is the lack of invariance problem. Listeners achieve phonetic constancy despite overlapping acoustic patterns of phonemes; few boundaries between phonemes or words; boundaries that shift with speaking rate, talker characteristics, phonetic context, coarticulation, and novelty of message content; and the ephemerality of elements of a spoken word in the environment. Speech recognition models are generally one of two extremes: complex, abiological real-word performant models that have done little to advance cognitive theories, and simpler, idealized models that have been theoretically influential but are severely limited in many respects. My collaborators and I have developed a model of human speech recognition, called EARSHOT, which lies between these two poles. EARSHOT shows good correspondence with human behavioral data and learns to recognize phonemes despite no explicit training on phonetic tasks. I discuss the EARSHOT model, and then its application to the dual lexicon hypothesis, which is the claim that the brain stores separate wordform representations for mapping to articulation (production) and meaning (semantics). We use EARSHOT to demonstrate that these separate representations may exist because of computational tradeoffs.
Dr. Kevin Brown is an associate professor in Pharmaceutical Sciences and Chemical, Biological, and Environmental Engineering at Oregon State University. He received his B.S. in physics and B. A. in mathematics from Louisiana State University and his PhD in theoretical physics from Cornell University. He was a Helen Hay Whitney Foundation fellow in Molecular and Cellular Biology at Harvard University, a postdoctoral fellow in Physics at the University of California, Santa Barbara, and an assistant professor in Biomedical Engineering prior to coming to OSU in 2018. He is a complex systems scientist who studies biological systems, particularly those arising in systems biology, systems neuroscience, and cognitive science. He is the originator of “Sloppy Models,” a theory of parameter space geometry in large nonlinear models with many underdetermined degrees of freedom. He has studied networks in molecular biology, neuroimaging, and cognitive science and employs a mix of data-driven and model-driven approaches. His work is tightly connected to experimental data, and he has many productive collaborations with experimentalists.