Douglas O'Shaughnessy
INRS-Telecommun., 16 Place du Commerce, Verdun, PQ H3E 1H6, Canada
Most analysis of speech examines carefully read texts. Natural spontaneous speech differs in several ways, having interruptions and large changes in speaking rate. The acoustic phenomena studied here concern the global speaking rate of spontaneous speech, how it varies for both fluent and disfluent speech. Such a model should find application in automatic speech synthesis and recognition. There is a tendency for speakers to adopt a specific articulation rate in fluent speech, especially for brief monosyllabic function words; stressed content words show greater variability, especially in prepausal situations. The greatest variability, however, occurs in disfluent speech, which have extremes in slow and fast speech. A large amount of the variability among word durations can be accounted for by: the number of phonemes per word, whether it is a function word or a content word, and whether the word forms part of a common sequence of words. Words at the start of a syntactic unit tended to be shorter than those at the end of the unit. The first time a content word was introduced into a discourse, it was stressed; repeated occurrences of such a word in the ensuing discourse had shortened durations. [Work supported by NSERC-Canada.]