A statistical approach to the segmentation and broad classification of continuous speech into phrase-sized information units

An algorithm is presented which uses the F₀ tracings of a connected-speech utterance as input and performs speaker-independent segmentation into prosodically defined information units. Two global declination lines are computed by the linear regression method, which approximate the trends in time of the peaks (topline) and valleys (baseline) of F₀ across the utterance. Computation is reiterated every time the Pearson product moment correlation coefficient for these declination lines drops below the present level of acceptability. Segmentation is thus performed without prior knowledge of higher level linguistic information, with the termination of one unit being determined by the general resetting of the intonation contour wherever in the utterance it may occur. The structure of the algorithm is described and its performance evaluated on three medium-sized Swedish texts read by four native speakers of standard Swedish