Data labeling method and device for mixing Chinese and English and labeling the tone

Abstract

The embodiment of the present application discloses a data labeling method and device for mixing Chinese and English and labeling the tone, which is applied in the deep learning speech synthesis algorithm. The method includes: grabbing the training text from the data source, and the training The text covers Chinese and English characters; add emotion tags to the captured training text, record the speaker's reading audio file of the training text marked according to the emotion tag, as the audio file for training; check the audio file for training with the corresponding Whether the emotional labels of the training text are consistent, and the audio files are revised for the inconsistent parts; the training text is mapped to a text vector, and the text vector and the audio file read by the speaker are submitted to the deep learning engine of the neural network for training. Learning and training, learning the pronunciation characteristics of the text under various combinations of Chinese, English, and emotional labels.

Description

technical field [0001] The embodiments of the present application relate to a data labeling method and device for mixing Chinese and English and labeling the tone. Background technique [0002] The current speech synthesis technology has greatly improved the quality of speech synthesis, and can directly generate realistic speech from text, which can be applied to fields such as voice navigation, automatic broadcast, and automatic queuing and calling services. However, in the current text-based voice output technology, the pitch is often flattened during the voice output process. Although it sounds smooth, it lacks emotional color and gives a very bad experience. At the same time, the traditional voice output technology cannot be applied to the mixed situation of Chinese and English at the same time. When it comes to mixed pronunciation of Chinese and English, two models are often called for processing, resulting in low processing efficiency and poor voice output effect. Th...

AI Technical Summary

Problems solved by technology

Under this kind of labeled data, due to the singleness of data preparation, it is basically impossible to train cadenced speech

Images