Voice recognition and voice synthesis model training method based on dual learning

Abstract

The invention provides a voice recognition and voice synthesis model training method based on dual learning. The method comprises the following steps that firstly voice recognition acts as the "main task" and voice synthesis acts as the "dual task"; the voice data A are converted into a text B' by using a voice recognition model to be trained; the confidence coefficient that the text obtained by conversion of the voice data A is written by humans rather than by machines is calculated by using a text language model obtained by pre-training; the text B' obtained by conversion of the voice data Ais converted back to the voice data A' by using a voice synthesis model to be trained; the "reconstruction similarity" between the voice data A' and the original voice data A is calculated by using avoice language model obtained by pre-training; and the final "reward" is calculated, and the parameters of the voice recognition model to be trained and the voice synthesis model to be trained are updated by using the REINFORCE algorithm of the reinforcement learning technology. A lot of cost overhead caused by data collection can be saved.

Description

technical field [0001] The present invention relates to the technical field of speech recognition and speech synthesis, in particular, it is a kind of speech recognition and speech synthesis established by using deep learning technology in an unsupervised manner by utilizing the nature of dual learning, using a large amount of unlabeled data and reinforcement learning technology. The method for training the speech synthesis model can be applied to the fields of speech recognition and speech synthesis. Background technique [0002] Speech is the most basic and most effective way for people to communicate in daily life. With the maturity of artificial intelligence technology, people also hope to communicate and transmit information with computers through direct dialogue, so speech recognition and speech synthesis have also become a major topic in the field of natural language processing. The demand for various forms such as speech-to-text and text-to-speech synthesis is expan...

AI Technical Summary

Problems solved by technology

[0003] Traditional speech recognition and speech synthesis model training methods require a large amount of one-to-one correspondence between speech data and text data, but collecting a sufficient amount of such one-to-one correspondence data is not only a time-consuming and laborious task, but also difficult to guarantee The quality of the collected data, in addition, the collection of data will also bring a lot of cost overhead

Insufficient amount of high-quality data has become a major obstacle to improving the accuracy and conversion efficiency of speech recognition and speech synthesis models

Images