Voice noise method and system for data enhancement

Abstract

The embodiment of the invention provides a voice noise adding method for data enhancement. The voice noise adding method includes the steps that speaker vectors of noise-free voice frequency and voicefrequency with noise are input into a conditional variational self-coding model, the vector mean and a variance vector outputted by a model encoder are sampled in a random gaussian distribution mode,and a noise implicit vector is obtained; the noise implicit vector and the noise-free voice frequency are input into the model, and simulation noise voice frequency is output by the model decoder; the model is based on the training conditions of the simulation noise voice frequency and the voice frequency with the noise, obtained varied noise implicit vectors are modeled, and the noise implicit variable space is obtained; and the noise implicit variable space is randomly sampled to be used as a noise adding implicit vector, the noise adding implicit vector and the voice frequency with the noise are input into the model decoder, and new voice frequency with the noise for data enhancement is obtained. The embodiment of the invention further provides a voice noise adding system for data enhancement. According to the voice noise adding method and system for data enhancement, the speaker vectors are modeled, through extraction of implicit space features, more varied noise data can be generated, and robustness of a voice recognition model is improved.

Description

technical field [0001] The invention relates to the field of voice recognition, in particular to a voice noise adding method and system for data enhancement. Background technique [0002] Over time, speech recognition technology has made great improvements, but when speech recognition is applied to environments with complex noise, it still has a certain impact on its recognition performance. In order to make the speech recognition model better applicable to various noise environments and improve the robustness and generalization ability of the speech recognition model to noise, it is usually trained with more noisy frequencies, because this way It is simple and effective, but it is difficult to obtain the noisy frequency suitable for training. For this reason, artificial noise is usually added to clean speech. For example, artificially collecting noise and then mixing the noise with clean speech yields more noisy frequencies suitable for training. [0003] In the process o...

AI Technical Summary

Problems solved by technology

[0005] In order to at least solve the problem in the prior art that due to the limited types and quantities of artificially added noise, the trained speech recognition model has poor robustness and generalization ability, and at the same time, the cost of artificially added noise is relatively high.

Images