An Analysis of Sentence Segmentation Features for Broadcast News, Broadcast Conversations, and Meetings

Information retrieval techniques for speech are based on those developed for text, and thus expect structured data as input. An essential task is to add sentence boundary information to the otherwise unannotated stream of words output by automatic speech recognition systems. We analyze sentence segmentation performance as a function of feature types and transcription (manual versus automatic) for news speech, meetings, and a new corpus of broadcast conversations. Results show that: (1) overall, features for broadcast news transfer well to meetings and broadcast conversations; (2) pitch and energy features perform similarly across corpora, whereas other features (duration, pause, turn-based, and lexical) show di erences; (3) the e ect of speech recognition errors is remarkably stable over features types and corpora, with the exception of lexical features for meetings, and (4) broadcast conversations, a new type of data for speech technology, behave more like news speech than like meetings for this task. Implications for modeling of di erent speaking styles in speech segmentation are discussed. General Terms Prosodic Modeling, Sentence Segmentation