انجمن کامپیوتر ایران

paper, we have used Tilt intonation theory to represent pitch contour as symbolic elements. Also, classification and regression trees are considered to estimate Tilt parameters in our text-to-speech system. To train regression trees, about 70 minutes of speech are used as corpus and more than 100 input features are extracted from this corpus. Further, some experiments such as incremental adding tilt parameters as input features during training and test models or purposing stress syllables as accent events are examined to achieve optimal regression trees. Moreover, vector quantization method and building a codebook of tilt parameters are investigated to predict pitch contours. Finally 61.1% and 25/386 hertz are obtained for correlation coefficient and RMSE values between predicted and real pitch contours of test set using optimal regression trees.