Prosodic prediction model training method, prosody prediction method and related device

Abstract

An embodiment of the present invention provides a prosody prediction model training method, a prosody prediction method and related devices. The training method includes: using the prosody prediction model to be trained to determine the current text unit and the previous text unit; obtaining the training current text unit of the current text unit The prediction vector is to obtain the previous prosodic prediction vector of the previous text unit, and fuse the two to obtain the training prosody fusion prediction vector; obtain the first training prediction prosodic vector according to the training current text prediction vector, and obtain the second training prosody fusion prediction vector according to the training prosody fusion prediction vector 2. Training and predicting prosodic vectors: obtaining the prediction loss of the current text unit according to the first training predicting prosody vector, the second training predicting prosody vector and the current reference prosody vector, adjusting the parameters of the model, and obtaining the trained prosody prediction model. The prosody prediction model training method, prosody prediction method and related devices provided in the embodiments of the present invention can improve the accuracy of prosody prediction.

Description

technical field [0001] The embodiments of the present invention relate to the field of computers, and in particular, to a prosody prediction model training method, a prosody prediction method and related devices. Background technique [0002] With the development of computer technology and deep learning technology, speech synthesis technology has become an important research direction and has been widely used, such as: voice broadcast, voice navigation and smart speakers. In speech synthesis, it is necessary to predict the prosody of the text, and the quality of the prosody prediction directly affects the naturalness of the synthesized speech and the accuracy of meaning expression at the semantic level of the text. [0003] When performing prosodic prediction, it is necessary to extract feature information from text information, including shallow feature information and deep feature information. The shallow feature information includes feature information such as word length...

AI Technical Summary

Problems solved by technology

[0003] When performing prosodic prediction, it is necessary to extract feature information from text information, including shallow feature information and deep feature information. The shallow feature information includes feature information such as word length, part of speech, and punctuation. , can be intuitively extracted from the text analysis results, but because the language information it covers is relatively superficial, it cannot reflect the semantic relationship between the texts in the sentence, and it is easy to cause the predicted prosody to be inaccurate, and the semantics of the prosodic unit cannot be guaranteed Integrity; for deep feature information, the information covered by each deep feature information is single, resulting in limited ability to use any kind of deep feature information for prosodic prediction. When combining multiple deep feature information, different There will be side effects between deep feature information, which will affect the accuracy of prosody prediction

Images