Music Classification based on Genre using LSTM

Music is one of the most popular entertainment media. The music industry has seen substantial growth in the past decade in streaming and releases. With streaming services, people can listen to music from all parts of the world. This increase in music demand makes music organizing a necessity for better access and recommendations. Genre is the most widely used characteristic feature of music to label them. In this paper, a dataset consisting of music from ten different genres is used to train a model that recognizes the music genre. Music tracks are processed to extract Mel Frequency Cepstral Coefficient features (MFCC). There are two approaches used to build a model to classify music. The first approach is to build an LSTM (Long Short Term Memory) network which is known to process time-series data and to recognize unique patterns for each class over time. The second approach uses a technique called transfer learning. A pre-trained LSTM network is used to extract embedding for each music file called d-vectors. The d-vectors are used to classify music using an SVM model.