A rule based perceptual intonation model for Turkish text-to-speech synthesis

In this study, a rule based perceptual intonation model was proposed for Turkish text-to-speech synthesis systems. A proper intonation model is required for a natural synthesized speech. The proposed model is sentence based and includes the stress patterns of compounds, inflected verbs and some punctuation marks. The offered rules were determined with the help of an interface where acoustical properties such as duration, pitch frequency and energy of 16 kHz sampled diphones could be adjusted and the rules were evaluated by the subjective Comparative Mean Opinion Score (CMOS) test. In the end, related with the quality of the synthesis, we can say that, the intonation model applied synthesis were found 1.38/5.00 points superior than raw synthesis and this shows the success of our proposed intonation model. Studies are being carried on expanding the test set.