Title
Statistical learning for integrating linguistic insights in automatic and human speech recognition models
Speaker
Eric Fosler-Lussier, The Ohio State University
Abstract
Speech recognition engineers have long used statistical learning techniques to model and transcribe spoken utterances. However, many current models are impoverished with respect to insights provided by the research of linguists. This talk will be in three parts: first, I will give a general overview of speech recognition as currently practiced. The second part of the talk will describe current research in the CSE Speech and Language Technology (SLaTe) lab which integrates posterior estimates of phonological features (subparts of speech sounds) in a Conditional Random Fields paradigm. The final, brief, part of the talk will discuss some recent in-progress work in modeling child language acquisition, where we employ self-organizing maps and Hebbian learning to model how kids acquire vowel spaces.
Joint work with Jeremy Morris, Ilana Heintz, Mary Beckman and Lucie Menard.
Meet the speaker in Room 212 Cockins Hall at 4:30 p.m. Refreshments will be served.
