基于双重数据增强策略的音频分类方法

Audio classification using double data augmentation strategy

卷积神经网络模型作为音频特征提取器具有较好的应用效果,但该类模型的训练过程对数据量要求比较高。针对这一问题,本文提出一种基于双重数据增强策略的音频分类方法。首先采用传统音频数据增强方法(旋转、调音、变调、加噪),并将增强后的数据转化为语谱图,再采用随机均值替换法进行谱图增强。在此基础上训练Inception_Resnet_V2神经网络模型作为音频特征提取器,最后训练随机森林模型作为分类器完成音频分类任务。实验结果表明,与已有方法相比,采用双重数据增强策略可明显提升音频分类精度,并且训练出的特征提取模型具有较强的泛化能力。

Convolutional neural network(CNN)model has been proved to be very effective in extracting audio features,but its training process requires a large amount of data.To solve this problem,an audio classification method using double data augmentation strategy is proposed.Firstly,the traditional audio data expansion approaches(rotation,tuning,frequency modulation and adding noise)are adopted and then the expanded data are transformed into spectrogram.The next step is to enhance the spectrogram by random mean replacement.On the basis of doubly enhanced data,Inception_Resnet_V2 neural network model is trained as audio feature extractor.Finally,the random forest model is trained as classifier to complete the audio classification.The experimental results show that,compared with the existing methods,the double data augmentation strategy can significantly improve the accuracy of audio classification,and the trained model for feature extraction has strong generalization ability.