Polyphone Disambiguation Based on Maximum Entropy Model in Mandarin Grapheme-to-Phoneme Conversion

Abstract

Polyphone disambiguation is the core issue of the grapheme-to-phoneme (G2P) conversion in Mandarin Text-to-Speech (TTS) system. In this paper, we propose a maximum entropy (ME) model to disambiguate polyphones, and evaluate various keyword selection approaches in different domains. Furthermore, we design a hierarchical clustering algorithm for automatic generation of feature templates, which minimizes the need for human supervision during ME model training. Results of comparative experiments show that, for the task of polyphone disambiguation, ME model obviously outperforms decision tree (DT), log-likelihood ratio is the best scoring measure of keyword selection, compared to manual templates, templates automatically generated by our hierarchical clustering algorithm significantly improve the accuracy of polyphone disambiguation, and greatly reduce the size of the ME model.