Sixth International Conference on Spoken Language Processing (ICSLP 2000)

Beijing, China
October 16-20, 2000

Syllable Recognition Using Glides Based on a Non-Linear Transformation

Jun Toyama, Masaru Shimbo

Graduate School of Engineering, Hokkaido University, Japan

The steady part of a phoneme becomes short or is sometimes even lost as its utterance speed increases or in natural conversation. Thus, a robust recognition model can be constructed by focusing on the glide passage parts in speech instead of on the phonemic stationary parts. On the other hand, the trajectories in feature space are generally curved. If an increase in likelihood corresponds to an increase in the degree of recognition accuracy, it is desirable that the curve-like trajectories are transformed into straight lines. In addition, an increase in the degree of recognition accuracy can be expected since degeneration of the feature space can be avoided by imposing the constraint that the distance between the centers of distributions is large. The results of computational evaluation tests on CV syllable glides for 50 ms showed syllable recognition rate of 87.1%.


Full Paper

Bibliographic reference.  Toyama, Jun / Shimbo, Masaru (2000): "Syllable recognition using glides based on a non-linear transformation", In ICSLP-2000, vol.1, 321-324.