Prosodic word boundary detection using statistical modeling of moraic fundamental frequency contours and its use for continuous speech recognition

A new method for prosodic word boundary detection in continuous speech was developed based on the statistical modeling of moraic transitions of fundamental frequency (F₀) contours, formerly proposed by the authors. In the developed method, F₀ contours of prosodic words were modeled separately according to the accent types. An input utterance was matched against the models and was divided into constituent prosodic words. By doing so, prosodic word boundaries can be obtained. The method was first applied to the boundary detection experiments of the ATR continuous speech corpus. With mora boundary locations given in the corpus, total detection rate reached 91.5%. Then the method was integrated into a continuous speech recognition scheme with unlimited vocabulary. A few percentage improvement was observed in mora recognition for the above corpus. Although all the experiments were done in closed conditions due to the corpus availability, the results indicated the usefulness of the proposed method