A Comprehensive Machine Learning Approach for Predicting Beats Per Minute (BPM) in Music Using Audio Features

Abstract

Predicting Beats Per Minute (BPM) in music is a significant challenge due to the complexity of the relationship between various audio features, such as rhythm, energy, and mood. Traditional methods are often unable to handle the complexity of feature variations and interactions. This study aims to develop a more accurate and reliable machine learning model to predict song BPM based on extracted audio features. We use advanced machine learning algorithms, including LightGBM, XGBoost, and Random Forest, to train models with a dataset covering ten audio features. Evaluation is performed using a k-fold cross-validation scheme with RMSE, MAE, and R² Score metrics. The experimental results show that boosting-based models such as LightGBM produce the best performance, with the lowest RMSE of 10.48, the lowest MAE of 7.62, and the highest R² Score of 0.83. However, these models still show a tendency to regress to the mean, indicating that some more extreme BPM variations are not fully captured. These findings emphasize the importance of improvements in feature engineering techniques and data rebalancing to improve BPM prediction accuracy in practical applications, such as music recommendation systems and tempo analysis.