A Speech Data Processing Method Based on Variational Gaussian Regression Process

Abstract

The invention discloses a speech data processing method based on a variational Gaussian regression process, and performs speech data processing based on a variational Gaussian regression process model. Transformation is carried out to make the final lower limit obtained larger, and the free variational Gaussian distribution of the active point set is obtained. The free variational Gaussian distribution is the posterior distribution of the selected points required in the mapping approximation, and the mapping approximation method is used to calculate Voice data is processed. The present invention improves calculation efficiency, makes approximations to some items in the finally obtained lower limit maximum value, and greatly improves calculation efficiency while minimizing similarity loss.

Description

technical field [0001] The invention relates to the field of voice data processing, in particular to a voice data processing method based on a variational Gaussian regression process Background technique [0002] The processing of speech data is a very important part of modern information data processing. By dividing frames according to time intervals, for each frame of speech data, it can be represented by a high-dimensional column vector, and each dimension corresponds to a feature, such as resonance Peak frequency, bandwidth, etc. In this way, the processing problem of speech data can be converted into a more common data processing problem, and then converted into a regression problem; [0003] Gaussian process regression is a machine learning regression method. It is a non-parametric regression method, so compared with the parametric regression method, its overfitting phenomenon is not serious, and the prediction results have probabilistic significance. At the same tim...

AI Technical Summary

Problems solved by technology

[0004] However, Gaussian process regression also has some problems. The most important thing is that the calculation cost is extremely high. In the current era, data processing involves huge data, so the standard Gaussian process regression is not used in practical applications. not promoted

In the existing technology, the biggest problem with the most standard Gaussian process regression is that the amount of calculation is too large, in other words, the calculation time is too long

The subsequent various approximations, including the VFE approximation, are all based on ensuring the accuracy as much as possible, so that the calculation amount is reduced, and the accuracy will inevitably be reduced.

Images