Lexical stress and speech recognition

Abstract

The characteristics of lexical stress and their possible use for automatic speech recognition are analyzed in this paper, with reference to the Italian spoken language. A theoretical analysis of the constraints imposed by the stress in a large lexicon access strategy was conducted on the basis of a 12 000 word vocabulary. Additionally, an investigation on several thousands of words spoken by different speakers was made in order to extract the statistical properties of the main stress correlates. An algorithm based on a statistical model succeeds to detect with a good reliability the stressed vowel of an isolated utterance. In the continuous speech case the same model can be used to compute a measure of the likelihood that a vowel is stressed. In the first case, the hypothesized stress position could be used as a source of constraints in a large lexicon access problem. In the second case, efficient pruning strategies based on highly reliable stress decisions can be designed for continuous speech understanding systems.