A Mandarin intonation prediction model that can output real pitch patterns

In this paper we proposed an intonation prediction model for Mandarin TTS systems. Our model can output real pitch patterns by finding a suitable real pitch pattern from the training corpus. This method is a new experiment. The advantages of our model are as follows. (1) It can improve the naturalness of the synthesized speech. It gets higher scores in the subjective listening tests. (2) It has high accuracies. Average errors of 0.425 ms and 0.457 ms were obtained for the inside and outside tests, respectively. Pattern errors of 0.128 ms and 0.129 ms were obtained for the inside and outside tests, respectively. We found that the pattern error measurement method complies with human hearing. (3) The training corpus need not be very large. It can relieve the data sparsity problem.