Ein quantitatives Tonmodell für Ibibio

In this Master's Thesis, a simple generic model for tone language synthesis is proposed and applied to an Ibibio corpus. While most parametric models (like Tilt and PaIntE) were designed for accent languages, or allowed for multiple solutions of their parameters (Fujisaki), the model proposed here consumes arbitrary tonal shapes while considering the tonological rules of the language. Additionally a software intonation module was developed to apply this model to the Bonn Open Synthesis System BOSS. Starting from theoretical considerations, tone contours are stylized syllable by syllable through polynomials. The whole corpus of those polynomial shapes is reduced by a simple vector quantization algorithm to a codebook of e.g. 64 codevectors called tone templates. By using about 20 common and easily extractable parameters from the foregoing modules, a classification tree predicts the likeliest tone template for each syllable. A standard unit selection algorithm selects segmental units with the best fitting contour shapes from the corpus. Thus the synthesis system aims to produce lexically and grammatically correct speech intonation. In this work, the problems and details leading to design decisions are discussed, the accuracy of the stylization, vector quantization and prediction are presented. Some proposals are made for further development of the model and software. The accompanying programming lead to the BOSS-IBB extension version 0.1 currently being integrated as module package version 0.2 into BOSS version 3.1.