LSTM and CNN-Based Detection of AI-Generated Classical Music From MIDI Features

Abstract

Detecting AI-generated classical music is a growing challenge as artificial intelligence continues to improve its ability to compose pieces that closely resemble human compositions. This study explores the use of deep learning methods like LSTM and CNN to classify whether a set of classical music is generated by AI or humans. The classification is based on sequential features extracted from MIDI files using beat-based segmentation, capturing statistical data of pitch, velocity, and duration over segments. The model was trained on a dataset comprising both AI-generated and human compositions, incorporating fine-tuning for optimal performance. Experimental results demonstrate that the proposed LSTM-based model achieves 99.00% accuracy on the primary test set, with an additional evaluation on an auxiliary dataset yielding 98.70% accuracy, confirming its reliability and strong generalization ability. Meanwhile, the CNN-based model attains accuracy scores of 97.00% and 97.10% on the primary and auxiliary datasets, respectively. Evaluation using confusion matrices and classification reports further validate both models' effectiveness, showing minimal misclassification rates. These findings suggest that while both LSTM and CNN achieve high classification performance in detecting AI-generated classical music, LSTM outperforms CNN in classification accuracy. Future research could explore integrating additional musical features or testing the model by expanding the dataset to cover a broader range of compositions, further improving model robustness and applicability.