Segmental durations in connected speech signals: preliminary results

The data base, methods for a study of the durations of phonetic units in connected speech, and some preliminary results are described. From readings of two scripts by many talkers, two sets of seven talkers each were selected, based on total reading time, to form a fast group and a slow group of talkers. Using computer graphics and digital playback procedures, the recordings were segmented into breath groups and pauses, and the first four sentences in each script were segmented into phones. The hold and release (that is, plosion and/or frication) portions of stops were identified and measured; less than 50% of the stops included releases. To establish the usefulness of the data base, the first-order statistics of the phonetic segments were determined, and a variety of durational characteristics were compared to existing reports. Analysis of number of breath groups, phonation time, and pause characterized the difference between so-called average fast and average slow talkers; however, no script-independent measure of these variables was found which would accurately predict the classification of individual talkers. The mean durations of various phonetic categories showed essentially the same percentage change when the fast and slow talkers were compared. Preliminary analyses of contextual influences on durations showed some expected changes, and also indicated that certain traditional predictions may not hold for informal connected speech. Gamma functions were fitted to the distributions of durations of various gross categories.