This paper presents our new approach to model tone coarticulation of Chinese continuous speech for tone recognition. We suggest that coarticulation effects between two neighboring tones are rather unstable, since they may be uni-directional, bi-directional, or none despite of the same phonetic contexts. Instability is suggested due to non-local prosodic events like prosodic phrase boundaries or stress effects. Hence, we propose that context dependent tone models should be estimated according to the exact underlying coarticulation effects. To simplify label work for coarticulation effects, label sets as few as 3 labels are adopted. Also F0 contours of tone nuclei are used to facilitate human to discriminate tone coarticulation effects. A new search algorithm for the output candidates was also proposed to adopt the new modeling method of tone coarticulation effects. Preliminary experiments on a females utterances of data corpus HKU96 showed the effectiveness of the new approach.
Cite as: Zhang, J.-S., Kawanami, H. (1999) Modeling carryover and anticipation effects for Chinese tone recognition. Proc. 6th European Conference on Speech Communication and Technology (Eurospeech 1999), 747-750, doi: 10.21437/Eurospeech.1999-181
@inproceedings{zhang99_eurospeech,
author={Jin-Song Zhang and Hiromichi Kawanami},
title={{Modeling carryover and anticipation effects for Chinese tone recognition}},
year=1999,
booktitle={Proc. 6th European Conference on Speech Communication and Technology (Eurospeech 1999)},
pages={747--750},
doi={10.21437/Eurospeech.1999-181}
}