Polyphone pronunciation discriminating method and device based on deep neural network

Abstract

The invention provides a polyphone pronunciation discriminating method and device based on a deep neural network. The method includes: quantitatively coding each Chinese character in a to-be-recognized text to generate feature vectors of the Chinese characters; according to the feature vectors of the Chinese characters, combining with bidirectional context information of the Chinese characters togenerate input features of the Chinese characters; respectively inputting the input features of the Chinese characters into DNN models corresponding to initial consonant, vowel and tone to respectively acquire first probability, second probability and third probability; calculating probability of various combinations of the consonant, the vowel and the tone according to the first probability, thesecond probability and the third probability, and taking the combinations with highest probability as pronunciation of the Chinese characters. By the method, pronunciation discriminating accuracy canbe improved, the problem that training samples are unbalanced caused by high-frequency sound of polyphones is effectively relieved, the pronunciation discriminating problem of multiple polyphones canbe solved, and integration of a voice synthesis system is facilitated.

Description

technical field [0001] The present application relates to the technical field of speech synthesis, and in particular to a method and device for discriminating the pronunciation of polyphonic characters based on a deep neural network. Background technique [0002] Word-to-sound conversion is an essential part of the Chinese speech synthesis system, and its accuracy directly affects the intelligibility of speech synthesis. Most Chinese characters have definite pronunciations, and the correct pronunciations can be obtained by searching a dictionary. However, there are many characters in Chinese that have two or more pronunciations. The key points and difficulties of word-sound conversion are the discrimination and disambiguation of these polyphonic characters. It is generally believed that the pronunciation of polyphonic characters is usually closely related to specific contextual information, semantics and language habits. How to automatically distinguish and analyze the pr...

AI Technical Summary

Problems solved by technology

[0006] However, the polyphone disambiguation method adopted in the existing related art relies on the contextual word segmentation information of limited distance, and the accuracy rate of polyphonic pronunciation discrimination is not high in texts modified by long attributives or long adverbials; That is, the pronunciation that often appears in the text), the model tends to predict a common pronunciation, and the prediction of the uncommon pronunciation is inaccurate; the rules or models obtained by training are only effective for a certain polyphonic character, and multiple polyphonic characters need to be trained more A model or rule, not suitable for training and pronunciation prediction of texts containing multiple polyphonic characters

Images