SUPRASEGMENTAL DURATION CONTROL WITH MATRIX PARSING IN CONTINUOUS SPEECH RECOGNITION

This paper describes a unified framework for continuous speech recogni tion (CSR) under grammatical constraints, where trellis calculations a nd parsing are performed by the same simple fundamental operations, na mely multiplication and addition of likelihood matrices. The matrix pa rser is shown to be a generalization of the CYK parser. It also facili tates explicit supra-segmental duration control for all grammatical ca tegories. Preliminary results showed that improved duration control on the mora level raised the recognition accuracy for a phrase recogniti on task from 86.7% to 88.5%.