Word boundary identification from phoneme sequence constraints in automatic continuous speech recognition

Abstract

This paper explores the extent to which phoneme sequence constraints can be used to identify word boundaries in continuous speech recognition. The input cousists of phonemic transcriptions (without word boundaries indicated) of 145 utterances produced by 1 RP speaker. The constraints are derived by matching the complete set of 3 phoneme sequences that can occur across word boundaries to entries in large lexicons containing both citation and reduced form pronunciations. Phonemic assimilatory adjustments across word boundaries are also taken into account. The results show that around 37% of all word boundaries can be correctly identified from a knowledge of such phoneme sequence contraints alone, and that this figure rises to 45% when a knowledge of one-and two-phoneme words and all legal, word-initial and word-final, two-phoneme sequences are taken into account. The possibility of including such constraints in the architecture of a continuous speech recogniser is discussed.