Method for synthesizing emotional speech by utilizing transfer learning under low resources

Abstract

The invention discloses a method for synthesizing emotional speech by using transfer learning under low resources, which comprises the following steps of: 1, pre-training an emotional vector: traininga speech emotion recognition model by using an EMOVDB data set, wherein the speech emotion recognition model is obtained by further processing a style vector extraction part in a basic method GST + Tacotron2 model of stylized end-to-end speech synthesis; 2, pre-training a speech synthesis model: for a basic Tacotron2 model, carrying out pre-training by utilizing a data set of LJSpeech 1.1; and 3,carrying out transfer learning training: for the basic Tacotron2 model, connecting the intermediate result obtained in the step 1 to the result of the encoder, and carrying out transfer learning training. According to the method, a pre-training and transfer learning method is adopted, a small amount of emotion data of a single speaker can be fully utilized, and on the basis of a unified emotion speech synthesis model, the synthesized speech with the quality reaching a certain level and the obvious emotion tendency is synthesized.

Description

technical field [0001] The invention relates to the field of speech synthesis, in particular to a method for implementing emotional speech synthesis by using existing data for migration learning under low resources. Background technique [0002] In recent years, the field of end-to-end speech synthesis has developed rapidly. Under the premise of training on large data sets, the quality and clarity of speech synthesis have been greatly improved. Emotional speech synthesis under the premise of training with a large amount of data has reached an acceptable level, but in some special cases, it may not be possible to obtain a dataset with a large amount of data for training, or the acquisition cost relatively high. SUMMARY OF THE INVENTION [0003] The purpose of the present invention is to overcome the deficiencies in the prior art, and to provide a method for emotional speech synthesis using migration learning under low resources. The methods of learning and model pre-train...

AI Technical Summary

Problems solved by technology

Emotional speech synthesis under the premise of training with a large amount of data has reached an acceptable level, but in some special cases, it may not be possible to obtain a large amount of data for training, or obtain a cost relatively high

Images