A Method of Polyphone Disambiguation Based on Semantic Extension
Qi Wang · 2021
In text-to-speech (TTS) system, polyphone disambiguation is the significant and difficult point in the process of grapheme-to-phoneme(G2P) conversion. In the existing methods, polyphone disambiguation mainly uses statistics or context information. When the data to be predicted is short texts or spoken language, it is difficult for the previous models to distinguish the correct pronunciation due to insufficient information. Therefore, this paper intends to report an improved design for polyphone disambiguation. Semantic extension has been used through the method of feature fusion, we creatively encode the context feature, word meaning feature, part of speech feature and pronunciation feature of polyphone, and construct three kinds of polyphone disambiguation models using bidirectional LSTM neural network. It is shown in the experiment that the accuracy of the polyphone disambiguation model based on word meaning can be improved by about 7% compared with the traditional methods.