Voice data enhancement method

Abstract

The invention discloses a voice data enhancement method and belongs to the technical field of voice data enhancement during machine learning processing. The method comprises steps that multiple automatic encoders are trained through the spectrogram of voice data in a training set, and the automatic encoders are separately trained through utilizing the spectrogram of a to-be-enhanced voice data set, so the N automatic encoders are obtained based on the to-be-enhanced voice data, the automatic encoders with these different structures are utilized to encode and represent original data, after theto-be-enhanced voice data spectrogram is inputted, multiple sets of structurally-differentiated output spectrogram structures are obtained, and lastly, the spectrogram structures are merged to obtaina newly-generated voice spectrogram data which can be trained, certain consistency is maintained compared with data before input in the main structure, but also some inconsistent expressions exist onsome structural features. The method is advantaged in that machine learning model performance based on the voice data is improved.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to the technical field of speech data enhancement during machine learning processing. Background technique [0002] One of the biggest difficulties in today's machine learning tasks is that it is difficult to train models when faced with small data sets. However, due to the particularity of some unnatural scenarios (business scenarios have strong timeliness, category labeling logic is difficult, and data label settings have a lot of subjectivity), it is also a very difficult task to collect and label data. Work. From the perspective of more common deep learning tasks such as image and text analysis, its data collection is relatively feasible to download, process and label millions of pictures and documents from the Internet. However, for speech data, not only is it difficult to collect and preprocess it, but also for a randomly collected speech, it is resource-intensive a...

AI Technical Summary

Problems solved by technology

[0008] The purpose of the present invention is to provide a data enhancement method for machine learning models in the speech category in view of the above-mentioned problems. The method can ensure that the machine learning model can train and analyze the spectrogram of speech. The original training data is used to synthesize the data, so that the amount of data and the form of data can be expanded on the basis of the original training data. In this way, the amount of data can be increased so that the machine learning model can be fully trained, and the user can You can try to use more complex machine learning models to fit speech-related problems, without the constraints and limitations between the amount of data and the amount of model parameters

Images