Segmentation-free syllable recognition in continuously spoken Japanese

A new recognition algorithm for continuously spoken Japanese is proposed. Recognition is accomplished by finding an optimal syllable sequence using dynamic programming (DP) to avoid the segmentation problem. CV (consonant-vowel) and VCV (vowel-consonant-vowel) templates are used to deal with the co-articulation problem. The VCV templates are prepared only for a subset of vowel and consonant combinations, as a description of the vowel-to-consonant transition, to reduce template registration labor as well as computational requirements. Template matching is made on a VCV level, also using dynamic programming, by connecting the VC part of the VCV template and the whole CV template. A 87% syllable recognition score was obtained for a male speaker. As an application example, large vocabulary word recognition was made. A 93% word recognition score was obtained for a 4300 word vocabulary.