Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

A Stochastic Polynomial Tone Model for Continuous Mandarin Speech

Yang Cao, Taiyi Huang, Bo Xu, Chengrong Li

National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China

In this paper, a stochastic polynomial tone model is presented for tone modeling in continuous mandarin speech. In this model, the pitch contour is described by a stochastic trajectory. The mean trajectory is represented by a polynomial function of normalized time while the variance is time varying. After that, an effective training and recognition algorithm is developed respectively. Also the problem of missing observation is discussed. Decision tree is employed to cluster the tone pattern variations, which are represented by proposed model. Many possible factors other than tone of neighboring syllables were taken into consideration when the decision tree was constructed. The experiments result shows that the tone recognition speed can increase more than 10 times while the recognition error rates decreased by 16% compared with traditional HMM tone model.

Full Paper

Bibliographic reference.  Cao, Yang / Huang, Taiyi / Xu, Bo / Li, Chengrong (2000): "A stochastic polynomial tone model for continuous Mandarin speech", In ICSLP-2000, vol.3, 674-677.