Incorporating linguistic theories of pronunciation variation into speech-recognition models

摘要

This paper describes the use of distinctive linguistic features to represent acoustic variability of words for speech recognition. Focusing on conventional hidden Markov model technology, we review implicit use of linguistic features as questions in decision-tree design for both coarticulation and pronunciation modelling and describe possibilities for more explicit use. The importance of conditioning on (hierarchical) syllable and prosodic structure is discussed, and the problem of modelling relative timing of feature-dependent acoustic cues is raised as a key limitation of current models. [References: 42]