Speech recognition using autosegmental representation of phonological units with interface to the trended HMM

Abstract

A novel speech recognizer is described which capitalizes on multi-dimensional articulatory structures and incorporates key ideas from autosegmental phonology and articulatory phonology. The novelty has been in the design of the atomic units of speech so as to arrive at a unified and parsimonious way to account for the context-dependent behavior of speech acoustics. At the heart of the recognizer is a procedure developed to automatically convert a probabilistic overlap pattern over five articulatory feature dimensions into a finite-state automaton which serves as the phonological construct of the recognizer. The phonetic-interface component of the recognizer, based on the nonstationary-state hidden Markov model or the trended HMM, is also described. Some phonetic recognition results using the TIMIT database are reported.