Speech is highly variable, especially in spontaneous conversation (words and sounds are acoustically very different in different contexts). The variation is systematic and predictable, and easily handled by human listeners, but presents a challenge for automatic speech recognition and for current models of human speech perception and production. Our objective is to integrate knowledge-based processing with statistical processing for speech analysis.
Our research uses knowledge of structure-driven variability to integrate knowledge-governed and statistical approaches. We extend signal processing methods to unify knowledge of acoustic, linguistic, physiological and cognitive models into a knowledge-based model of lexical access, and test the system as a model of human speech perception by determining its robustness to contextual variation and its breakdown in the same ways as human speech processing.
Cognitive Processing Constraints
Current Statistical Pattern Matching Methods
The group is led by Dr. Stefanie Shattuck-Hufnagel