Research On Music Genre Classification Method Based On LSTM Model

Music genre classification is an essential and important task in the field of music information retrieval,and it is a necessary part in music recommendation and automatic music tagging.The classification of music genres is very useful for users to search for their favorite music.The classification of music genres has always been the core problem of understanding human's preference for music,and it has been applied to the construction of music recommendation system.The tasks studied in the classification of music genres mainly include feature extraction and classifier design,and the use of various audio signal feature extraction techniques,the use of advanced features such as loudness,rhythm and timbre,as well as the most advanced neural network model has been widely studied.In the traditional classification method of music genre based on frame feature,the audio data is represented by independent frame,which completely ignores the temproal characteristics of audio.If the sequence relationship can be well modeled,the classification performance will be significantly improved.It is interesting and challenging to design and implement the music genre classification system by using the deep learning model.In this paper,around the classification of music genre,from audio feature extraction,classifier training to the final music genre prediction design and implementation of a complete music genre automatic recognition system.We use a long-term memory(LSTM)model for the classification of music genres,rather than convolution neural networks.By training a deep model,we can categorize music in eight different genres.In addition,we have adopted a hierarchical classification scheme to further improve the accuracy.First of all,we divide the music into two categories,namely,strong genre and weak genre.The music is divided into these two categories by training the LSTM classifier.The music is then further divided into multiple subclasses.In this paper,a music genre classification algorithm based on LSTM is proposed,which uses LSTM to learn the representation of continuous frame features,and integrates segmented features to consist of statistical information of frame features in each segment.Experiments on the GTZAN dataset show that the proposed music genre classification system can achieve 50% accuracy on the ten genre,which is about 3.13% higher than the baseline model of CNN.This significant improvement shows the effectiveness of the proposed deep LSTM algorithm in the classification of music genres.