Chinese Tone Modeling with Stem-ML

Production studies of Chinese tones show that tone shapes often deviate from the expected canonical shapes in natural sentences. The situation is particularly difficult in conversation where a tone can be realized with a shape opposite to the underlying specification. We note that the unexpected tone shape is the property of weak prosodic strength, where a weak tone accommodates the shapes of neighboring strong tones. If the tonal trajectory linking neighboring strong tones is in contradiction to the tonal specification of a weak tone, the weak tone gives in. This phenomenon supports the design considerations behind Stem-ML (Soft TEMplate Markup Language), a prosodic markup language we designed for multilingual prosody modeling. Tones/accents of a language are represented as soft templates which bend to conform to their environment. This paper reports the result of fitting Chinese tones with Stem-ML. We used invariant tone templates to specified the basic shapes of lexical tones. The surface variations are accounted for by linguistically motivated variations in prosodic strength.