The validation of automatic methods for measuring conversational speech activity

Sophisticated electronic systems have been developed to measure speech activity patterns automatically, but their accuracy is unknown. The purpose of the present work is to evaluate the fidelity with which a class of computerized systems matches the measurements made by a human observer. With all parameters optimized, it was found that: (1) about 98% of samples (sampling rate = 200/sec) were classified the same way by the system and the criterion method; (2) distributions of utterance durations and speaker-switch intervals were accurately rendered by the automatic system; (3) average durations of talkspurts—bursts of speech activity within utterances—were closely approximated, but the system tended to overestimate the number of brief (?100 msec) talkspurts; (4) the system was subject to considerable qualitative error in the measurement of within-utterance pauses and of simultaneous talking.