Modeling the cross-linguistic variations of tonal systems

This study aims to simulate the cross-linguistic variations of tonal systems with low dimensional models. Individual syllables of Mandarin, Yoruba, southern Yi, and Hmong were retrieved from existing speech corpora. Voice quality measures as well as f0 were extracted for all data. For each language, Functional Principal Component Analysis (FPCA) is used to parameterize the f0 contours of individual syllables. It is found that three FPCs are sufficient to account for more than 90% of the variance for pitch contours for all languages. Moreover, the same FPCs are shared among tonal languages and appear to be phonologically meaningful: FPC1 is related to the pitch range; FPC2 is related to the direction of the contour (e.g. rising or falling); FPC3 is related to more complex contours such as dipping and convex. Tone classification models were built for each language, and both voice quality measures and f0 PCs were fitted into the models. Voice quality cues significantly improve the accuracy of the predictions, but different languages vary in the relative importance of voice quality cues. This study provides a new effective way to computationally model the cross-linguistic variation of tonal systems, and has practical implications for speech synthesis.