Music Audio Sentiment Classification Based on Improvied Vision Transformer

Abstract

Common neural network models have problems of low accuracy and low efficiency in music sentiment classification tasks. In order to further excavate sentiment information contained in the audio spectrum and increase the accuracy of music sentiment classification, an improved Vision Transformer model is proposed. Since the public data set does not meet the requirements of the task of music sentiment classification, this paper makes a four-category music sentiment data set. After the audio is preprocessed, the processed audio features are trained by Vision Transformer. Modify the input of Vision Transformer to fit the structure of Vision Transformer. Position parameters of Vision Transformer model can better preserve the connection between audio features. Encoder structure can also fully learn local features and global features. Due to the long training time of this model, softpool pooling layer is introduced into the model, which can better retain the emotional features, speed up the calculation of the model, but also retain the model accuracy. Experimental results show that the classification accuracy of Vision Transformer model reaches 86.5%, which has better classification effect compared with neural networks such as ResNet. Meanwhile, the improved Vision Transformer reduces training time by 10.4% and accuracy by only 0.3%. On the public data set gtzan, the accuracy of this model reaches 90.7%.