Deep-core-study-based Gaussian process regression modeling method in voice conversion

Abstract

The invention discloses a deep-core-study-based Gaussian process regression modeling method in voice conversion, which comprises the following steps of intercepting the first 2 / 3 of one-dimensional data in multi-dimensional data of tone features of source voice and multi-dimensional data of target voice features as training data; intercepting the last 1 / 3 data of multi-dimensional data of the tonefeatures of the source voice as test data; performing training through a Gaussian regression model; firstly, inputting the training data into a BP neural network to obtain an initial weight parameter; solving partial derivative on the weight parameter when hyper-parameter in a Gaussian process is unchanged; solving the partial derivative on hyper-parameter when the weight parameter is unchanged;then, updating the hyper-parameter and spreading the updated weight parameter by direction; obtaining a best result through loop iteration. The precision of the Gaussian regression model in original voice conversion is improved by the invention; a deep core and an expandable deep layer structure with better expression force is developed for combining the advantages of a neural network and the Gaussian process; all parameters are trained together through unified supervision, and are used as a part of a non-parameter Gaussian process frame; further breakthrough is realized on the expansibility and the depth of the Gaussian process.

Description

technical field [0001] The invention relates to a voice conversion based on deep kernel learning Gaussian process regression modeling method, belonging to the field of voice data processing. Background technique [0002] Speech data is an important processing content of modern information data. Each frame of speech data can be described by characteristic parameters. For example, the parameters related to formants are the formant frequency (first dimension) and bandwidth ( The second dimension), energy spectrum tilt (third dimension), etc., the high-dimensional data of the speech data frame described by the characteristic parameters, the high-dimensional data extracted from each frame can be reduced to three-dimensional, and regression and prediction can be performed; [0003] Gaussian process regression is a machine learning regression method. It has a strict theoretical basis for statistical learning. It has good applicability to complex problems such as high-dimensionality...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the problem of improving the accuracy of the Gaussian regression model in voice conversion, and to develop a more expressive deep kernel and scalable deep structure for the advantages of the joint neural network and Gaussian process, the present invention discloses a voice conversion based on a deep kernel Learning the Gaussian process regression modeling method, all parameters are trained together under unified supervision. As part of the non-parametric Gaussian process framework, the neural network provides a powerful learning machine for creating adaptive basis functions, rather than parametric models. Gaussian process provides great flexibility and automatic calibration

Images