Polyphone disambiguation and rhythm control combined method and system and electronic equipment

Abstract

The invention provides a polyphone disambiguation and rhythm control combined method, a polyphone disambiguation and rhythm control combined system and electronic equipment. The method comprises the steps of: obtaining a to-be-processed text and part-of-speech thereof, converting the to-be-processed text and the part-of-speech thereof into a character vector and a part-of-speech vector, and splicing to obtain a spliced vector; training through an alternate training strategy to obtain a joint model, a first group of weights and a second group of weights, wherein the joint model comprises a first neural network and a second neural network, and encoding the splicing vector through the joint model to obtain a first in-sentence code and a second in-sentence code of the character; obtaining a polyphone weighted sum according to the first group of weights, and obtaining pronunciation probability distribution of polyphones through a first full connection layer; removing incorrect pronunciation in the pronunciation probability distribution of the polyphone through masks, and obtaining final pronunciation prediction; and obtaining a rhythm weighted sum according to the second group of weights, and obtaining a rhythm pause level through a second full connection layer and a conditional random field. Error accumulation caused by stream structure processing is eliminated, and the calculation speed of text-to-speech conversion is improved.

Description

technical field [0001] The present application relates to the technical field of Chinese speech synthesis, and in particular to a polyphone disambiguation and prosody control joint method, system and electronic equipment. Background technique [0002] In order to avoid mispronunciation of polyphonic characters generated by text-to-speech technology or too bland speech, and to make the synthesized speech more accurate and "personalized", polyphonic word disambiguation and prosodic pauses to control speech are often added during the processing. [0003] In the traditional processing method, text-to-speech mainly includes two parts: front-end text / phoneme conversion processing and back-end phoneme / speech signal conversion processing. The processing of the back-end is based on acoustic features, which is used to achieve end-to-end training and synthesis; while the front-end includes the clause segmentation model, text regularization model, natural tone sandhi model, polyphone di...

AI Technical Summary

Problems solved by technology

[0004] At present, the general practice is to use two independent polyphone disambiguation models and prosodic prediction models in the front-end processing to realize polyphone disambiguation and prosodic prediction respectively. Structural processing leads to the accumulation of errors in each module, which affects the final speech synthesis effect

Images