Many-to-many speaker conversion method based on Perceptual STARGAN

Abstract

The invention discloses a many-to-many speaker conversion method based on perceptual STARGAN. The method comprises a training stage and a conversion stage. The STARGAN and a perception network are combined to realize a speech conversion system. The perception loss is calculated by using the perception network to improve the extraction capability of a model for the deep semantic features and personal features of a speech spectrum, so that the learning capability of the model for the semantics and personal features of the speech spectrum can be improved well, thereby improving the personal similarity and speech quality of speech generated after conversion well, solving the problem of poor similarity and natural degree of the speech generated after conversion in the STARGAN, and implementinga high-quality voice conversion method. Through adoption of the method, the voice conversion under a non-parallel text condition can be realized; no alignment process is required in the training process; a conversions system of a plurality of source-target speaker pairs can be integrated into one conversion model; the model complexity is lowered; and the multi-speaker to multi-speaker conversion is realized.

Description

technical field [0001] The invention relates to a many-to-many speaker conversion method, in particular to a perceptual STARGAN-based many-to-many speaker conversion method. Background technique [0002] Speech conversion is a research branch in the field of speech signal processing, which is developed and extended on the basis of speech analysis, recognition and synthesis. The goal of voice conversion is to change the voice personality of the source speaker so that it has the voice personality of the target speaker, while retaining semantic information, that is, to make the voice of the source speaker sound like the voice of the target speaker after conversion. . [0003] After years of research on voice conversion technology, many classic conversion methods have emerged. These include Gaussian Mixed Model (GMM), Neural networks (NNs) (including Restricted Boltzmann machines (RBMs), Feed forward NNs (FNNs), Recurrent NNs (RNNs), Convolutional NNs (CNNs), Long Short-Term ...

AI Technical Summary

Problems solved by technology

[0007] Purpose of the invention: the technical problem to be solved by the present invention is to provide a perceptual STARGAN-based many-to-many speaker conversion method, which solves the problem of network degradation in the training process of existing methods, reduces the learning difficulty of the coding network for semantics, Realize the learning function of the deep spectrum features of the model, improve the spectrum generation quality of the decoding network, and avoid the information loss and noise problems caused by the Batch Norm process, and more fully learn the semantic features and the speaker's personalized features, so as to compare Greatly improve the personality similarity and voice quality of the converted voice

Images