Speech synthesis method and device, electronic equipment and storage medium

Abstract

The invention discloses a speech synthesis method and device, electronic equipment and a storage medium, relates to the artificial intelligence technology, and uses the machine learning technology inartificial intelligence to carry out speech synthesis, and the method comprises the following steps: obtaining a word segmentation sequence corresponding to an input text; determining a vector sequence corresponding to the word segmentation sequence by utilizing a trained attention mechanism-based language representation model; carrying out rhythm structure prediction processing on the vector sequence by utilizing a rhythm structure model, and determining rhythm structure information which comprises pause duration and pronunciation intensity corresponding to word segmentation fragments corresponding to each feature vector in the vector sequence in the synthesized speech; and synthesizing speech corresponding to the input text according to the rhythm structure information and pronunciationcorresponding to each word segmentation fragment in the word segmentation sequence. According to the speech synthesis method and device, the electronic equipment and the storage medium provided by theinvention, the accurate and real semantics of the input text can be expressed based on the vector sequence obtained by the language representation model, so that the synthesized speech is more natural.

Description

technical field [0001] The present application relates to the technical field of artificial intelligence, in particular to a speech synthesis method, device, electronic equipment and storage medium. Background technique [0002] Speech synthesis technology (Text-to-Speech, TTS) is a technology that converts text input into speech output. It specifically involves multiple fields such as acoustics, linguistics, and digital signal processing multimedia technology. It has been widely used in human-computer interaction. [0003] At present, generally based on the RNN (Recurrent Neural Network, recurrent neural network) language model to generate a vector sequence that can represent the contextual semantics of the input text, and then perform prosodic structure prediction and pronunciation prediction based on the vector sequence, so as to predict the result based on the prosodic structure Synthesize speech with the pronunciation prediction result. [0004] However, the RNN langua...

AI Technical Summary

Problems solved by technology

[0004] However, the RNN language model has the problem of memory attenuation, and cannot learn the long-distance dependencies between words in a long input text, which leads to the fact that the vector sequence generated based on the RNN language model cannot represent the accurate and true semantics of the input text. Reduced the accuracy of subsequent prosodic structure prediction and pronunciation prediction, affecting the final synthesized speech effect

Images