A prosodic labeling method, device, equipment and medium

Abstract

The present application discloses a method, device, device, and medium for prosodic labeling. The method includes: acquiring first acoustic features, first text features, and first prosodic labeling results corresponding to sample audio; As the input of the encoder in the end-to-end neural network, the first text feature is used as the input of the decoder in the end-to-end neural network, and the first prosodic labeling result is used as the input of the end-to-end neural network Output, train the end-to-end neural network to obtain the end-to-end neural network after training; when the second acoustic feature and the second text feature of the prosody to be marked are obtained, use the end-to-end neural network after the training to directly output Second prosody labeling results. The prosodic labeling method effectively fuses the acoustic features with the corresponding text features, and improves the accuracy of prosodic labeling.

Description

technical field [0001] The present application relates to the technical field of speech synthesis, in particular to a prosody tagging method, device, equipment, and medium. Background technique [0002] The synthesized sound library generally includes a large number of high-quality recorded audio clips, corresponding transcribed texts, and prosodic annotations on the transcribed texts based on the prosodic information of the recorded audio clips. How to automatically and accurately carry out the prosodic labeling of the synthesized sound bank by computer has become an important technology in the field of speech synthesis. [0003] Existing technical solution 1: first use the pre-trained text prosody prediction model to predict the prosody information of the text, and then use the pre-recorded audio to authenticate and screen the predicted text prosody information, eliminate the incorrect prosody information, and keep the correct prosody information. prosodic information to ...

AI Technical Summary

Problems solved by technology

[0004] Solution 1 cannot finally filter out the correct prosodic information without reading according to the predicted prosodic rhythm of the text

Option 2 splits the internal connection between speech and text, and cannot achieve good prosodic labeling effect

Moreover, the prosodic labeling process in the existing solutions includes multiple stages of processing. The component construction of each stage requires rich domain knowledge. The design of the entire system is difficult and the implementation is complicated. The errors of each stage will be superimposed in the final stage, so that The resulting prosodic information is inaccurate

Images