Audio starting point labeling method and device thereof

Abstract

The invention discloses an audio starting point labeling method and a device thereof, and the method comprises the steps: obtaining an audio source file, carrying out the time-frequency analysis of the audio source file, and obtaining a spectrum energy diagram; carrying out smooth filtering on the spectrum energy diagram; establishing an oscillogram by taking the time domain in the audio source file as an abscissa and the amplitude as an ordinate, aligning the time domain of the abscissa of the oscillogram and the frequency spectrum energy diagram subjected to smooth filtering, and splicing upand down after alignment to obtain a comparison diagram for display; generating starting point labeling information according to the displayed comparison graph; preliminarily determining a pluralityof corresponding starting points in the audio source file according to the plurality of received time points, and judging and deleting an error starting point in the plurality of starting points; andlabeling the audio frame corresponding to the time point in the audio source file as a starting frame according to the finally determined time point corresponding to the starting point, and exportingthe starting frame as a labeling sample. The method has the advantages of high labeling accuracy and high generalization.

Description

technical field [0001] The present application belongs to the technical field of audio signal analysis and processing, and in particular relates to a method and device for marking an audio starting point. Background technique [0002] The note start point is the most basic feature in the music feature information, which refers to the time when a certain note in the music starts. Such as figure 1 As shown, in the time domain information of a note, the energy at the attack stage suddenly rises, and after a transition period (Transient), the energy gradually decreases (Decay). The onset of the attack stage is the note onset. starting point. Note onset detection has many application directions and important uses in the field of signal processing, such as: dividing music into beats, rhythm detection, pitch estimation, etc. [0003] Currently, there are two main types of labeling tools on the market, both of which use deep learning to extract audio features. It should be noted...

AI Technical Summary

Problems solved by technology

In this scheme, the value obtained by taking the logarithm of the short-time Fourier transform and the mel spectrum slice is used as the label. On the one hand, the length of the slice needs to be compared with the time domain one by one, and the workload is large; on the other hand, the training convolutional neural network The accuracy and precision of the network results cannot be effectively guaranteed. If there is a problem in training, it may not be effectively solved, and the generalization of the model to different types of audio signals is weak

Images