Accent type recognition and syntactic boundary detection of Japanese using statistical modeling of moraic transitions of fundamental frequency contours

Experiments on accent type recognition and syntactic boundary detection of Japanese speech were conducted based on the statistical modeling of voice fundamental frequency contours formerly proposed by the authors. In the proposed modeling, fundamental frequency contours are segmented into moraic units to generate moraic contours, which are further represented by discrete codes. After modeling the accent types and syntactic boundaries, their recognition/detection was done for ATR speech corpus. For the accent type recognition, 4-mora words were used for the training and testing, and recognition rates of around 74% were obtained for speaker open experiments. For syntactic boundary detection, the detectability of accent phrase boundaries was tested for sentence speech. Although the experiments were conducted only for the closed condition due to the availability of speech corpus, the result indicated the usefulness of separating the boundary model into two depending on whether the boundary is accompanied by a pause or not