Research On Music Genre Classification Algorithm Based On Attention Mechanism

With the development of mobile terminals and Internet technology,people have increasingly convenient mediums to obtain digital music.However,complex music genres and massive music libraries have brought great challenges to music information retrieval.Nowadays,various songs can always be linked due to some commonalities."Genre" is one of their commonalities.Music genres are high-level labels for music information,which would consume a lot of time and resources when manually tagged.At the same time,most music playback platforms construct music recommendation systems based on music genre classification models,which can provide recommendations to customers or just be launched as commercial products.In this research direction,recognizing genre of music is the first step.Thanks to the rapid development of computer hardware and acoustics,deep learning has been widely used in the study of music genre classification.However,the existing proposed machine learning models still have shortcomings such as low classification accuracy,difficult parameter tuning,and low practicability.In summary,in order to explore more effective models,this article proposes different algorithms from two ideas based on the attention mechanism: one is to transform training data into a combination of several audio features based on timing;the other is to transform training data into spectral images for training and classification.The specific implementation is mainly to combine different neural networks with a variety of different preprocessing technologies,feature extraction schemes,and optimization methods.At last,we applied our algorithms to the GTZAN which is the most classic dataset in this field to prove the reliability of the proposed models.Specifically,the main research contents of this article are as follows:(1)First,according to the temporal characteristics of music data,a classification model based on feature combination is proposed.The attention mechanism is designed based on the characteristics of the temporal classification model LSTM,thereby constructing a hybrid model including the Attention LSTM module and the onedimensional convolution module.The case results show that the proposed hybrid model can effectively improve the classification accuracy.(2)In order to further reduce the experimental loss,this paper draws on the design idea of the image classification model,converts the audio data into the spectrogram as the training data,and uses the advantages of the convolutional neural network in image processing.The classification model that combines the force mechanism with the CNN model optimizes the model's ability to learn feature maps.The experimental results show that the proposed model can not only improve the classification accuracy,but also has strong stability in training compared with the classification model based on feature combination.(3)In order to enhance the performance of the model,this paper uses Bayesian optimization as a hyperparameter tuning algorithm to find the optimal value of the attention module parameters,and finally achieved a result far exceeding manual tuning,and combined with multiple traditional machine learning algorithms,such as support vector machines,logistic regression,decision trees,random forests,and the latest deep learning models in the field for comparison,all have a greater advantage,can achieve higher classification accuracy,and have a wide range of research significance and application prospects.