Framework for Groove Rating in Exercise-Enhancing Music Based on a CNN–TCN Architecture with Integrated Entropy Regularization and Pooling

Groove, a complex aspect of music perception, plays a crucial role in eliciting emotional and physical responses from listeners. However, accurately quantifying and predicting groove remains challenging due to its intricate acoustic features. To address this, we propose a novel framework for groove rating that integrates Convolutional Neural Networks (CNNs) with Temporal Convolutional Networks (TCNs), enhanced by entropy regularization and entropy-pooling techniques. Our approach processes audio files into Mel-spectrograms, which are analyzed by a CNN for feature extraction and by a TCN to capture long-range temporal dependencies, enabling precise groove-level prediction. Experimental results show that our CNN-TCN framework significantly outperforms benchmark methods in predictive accuracy. The integration of entropy pooling and regularization is critical, with their omission leading to notable reductions in R² values. Our method also surpasses the performance of CNN and other machine-learning models, including long short-term memory (LSTM) networks and support vector machine (SVM) variants. This study establishes a strong foundation for the automated assessment of musical groove, with potential applications in music education, therapy, and composition. Future research will focus on expanding the dataset, enhancing model generalization, and exploring additional machine-learning techniques to further elucidate the factors influencing groove perception.