Система аналізу схожості мелодій для виявлення можливого плагіату

Abstract

This paper presents an approach to improving the effectiveness of musical plagiarism detection through the implementation of an intelligent information technology for melody similarity analysis. The integration of such a technology into the workflows of music producers, copyright specialists, forensic experts, and analysts in the music industry provides the ability to quickly and objectively assess the degree of similarity between musical works, reducing subjectivity and increasing the transparency of expert conclusions. As a result, users obtain a tool for rapid identification of potential cases of musical plagiarism based on short local melodic fragments.The system implementation is based on the combination of deep learning methods and convolutional neural networks, which are used to construct vector representations of melodic fragments. The data preprocessing module is responsible for tempo normalization, splitting melodies into fragments in MIDI format, and forming the piano-roll representation of a musical segment, which ensures the correctness of further analysis. During training, the convolutional neural network is optimized using a supervised contrastive approach that forms a feature space with a clear separation between similar and dissimilar fragments. The use of contrastive learningin combination with dividing compositions into short local fragments compensates for the limitations of traditional approaches and significantly improves the accuracy of detecting potential musical plagiarism.To test the proposed concept, a prototype of the system was developed as a software solution implemented in Python using the PyTorch, NumPy, pandas, and hnswlib libraries. The prototype performs the full cycle of MIDI fragment processing–from segmentation and vector representation construction to HNSW index formation and nearest neighbor search. The system provides batch processing of large datasets and maintains scalability during melody similarity search, enabling effective evaluation of the model’s performance in different plagiarism detection scenarios.The proposed approach creates the foundations for an intelligent system supporting copyright expertise, capable of automating the process of detecting potential musical plagiarism and increasing the objectivity of similarity assessment between works. In future research, the concept may be extended into a fully adaptive system capable of self-learning basedon accumulated musical data and gradually improving the accuracy of identifying potential copyright infringements.