Quantitative knowledge on word structure, from a phonetic corpus, with application to large vocabularies recognition systems

We have used hand-made phonetic transcriptions of 11 hours of spoken French, in order to study, with frequency measurements and elementary information theory, the strength of phonotactic constraints within isolated words, as opposed to continuous speech. Then, using a phonetic dictionary extracted from this corpus, we have studied the relative importance of gross phonetic classes oppositions, and of distinctive features.