Audio multi-label classification method based on deep learning

Abstract

The invention relates to the field of audio marking of environmental sound recognition, in particular to a multi-label classification method based on deep learning for noisy audio. According to the content of the invention, data preprocessing comprises performing noise reduction processing on a data set by using an RNNoise algorithm. The audio feature extraction comprises carrying out short-time Fourier transform on an audio, then converting the audio into MFCC feature data, and then inputting the MFCC feature data into a VGGish network to obtain 128-dimensional high level feature embedding; the model construction comprises the steps that a CNN and an RNN neural network are determined to be used, the CNN can well utilize a two-dimensional structure of input data to process voice data, andthe RNN can well utilize correlation between labels to orderly predict the labels; the model training comprises tracking a loss function value and a classification error, and updating model parametersuntil a model with relatively high accuracy is obtained. The model evaluation comprises defining evaluation indexes and calculating average precision; the audio multi-label classification comprises the steps of loading the trained model and outputting a predicted label probability result. The process is shown in Figure 1.

Description

technical field [0001] The invention relates to the field of audio marking for environmental sound recognition, in particular to a deep learning-based multi-label classification method for audio with noise. Specifically, after the audio feature is extracted, it is used as the input of the neural network for training to obtain a model with high accuracy, so as to perform label classification. Background technique [0002] In recent years, deep learning has been widely used in speech recognition, image classification, automatic driving and other fields, and the classification of environmental sound recognition is a problem that is widely used in real life. At present, the research on this problem is gradually became a hotspot. [0003] Traditional single-label classification mainly solves the problem that an example belongs to only one category. However, in real life, due to the complexity and polysemy of the objective object itself, there is often no absolute single-label c...

AI Technical Summary

Problems solved by technology

[0005] The invention discloses an audio multi-label classification method based on deep learning, which solves the problem of automatically classifying complex environmental sounds under noise interference

Images