Abstract: Grapheme-to-phoneme conversion(G2P) is an important part of the Chinese text-to-speech system(TTS).The key issue of G2P is to select the correct pronunciation for polyphonic characters among several alternatives.Existing methods usually struggle to fully grasp the semantics of words that contain polyphonic characters,and fail to effectively handle the imbalanced distribution in datasets.To solve these problems,this paper proposes a polyphone disambiguation method based on the pre-trained model RoBERTa,called cross-lingual translation RoBERTa(CLTRoBERTa).Firstly,the cross-lingual translation module gene-rates another translation of the word containing the polyphonic character as an additional input feature to improve the model’s semantic comprehension.Secondly,the hierarchical learning rate optimization strategy is employed to adapt the different layers of the neural network.Finally,the model is enhanced with the sample weight module to address the imbalanced distribution in the dataset.Experimental results show that CLTRoBERTa mitigates performance differences caused by uneven dataset distribution and achieves a 99.08% accuracy on the public Chinese polyphone with pinyin(CPP) dataset,outperforming other baseline models.

Key words: Polyphone disambiguation, Pre-trained model, Grapheme-to-phoneme conversion, Cross-lingual translation, Hierarchical learning rate, Sample weight