BiLSTM-Based Approach for Music Genre Classification

Due to the rise of artificial intelligence, recommending various types of music to meet users’ needs is becoming increasingly important. In the past, processing massive data sets to categorize music was an essential part of the work. Nowadays, many researchers have applied machine learning research methods to categorize music. These studies depend on the type of sample selection, some of which are not amenable to multiple categorizations or long computation times. Given this, our study will improve the sample selection method and train a better prediction model based on our sample data. In this paper, we use the GTZAN dataset as our music dataset, which contains ten different types of music; each type has 100 WAV audio files of about 30 seconds. To find the features of each category, we extract the Mel Frequency Cepstral Constant (MFCC) of each music file as a feature vector. Next, we used a Bi-Directional Long Short-Term Memory (BiLSTM) model to train the model. Our research results show that the accuracy of the BiLSTM model for 50 iterations is 80%; when we increase the number of iterations to 100, the accuracy of the trained model increases to 84%. Comparing the methods used by other research works, our model is suitable for long-term calculations. Although the execution time of BiLSTM is longer, the model’s performance will improve as the training time increases. In this research, we hope that people who design music recommender systems can have more considerable models to recommend better music that is closer to what the listener wants.