Music plagiarism detection using neural network models

Abstract

In recent years, the creation and publication of musical works have become significantly simplified thanks to the spread of digital technologies, generative artificial intelligence, and tools based on it. At the same time, detecting musical plagiarism remains crucial for copyrights holders, as the timely detection of plagia rized elements helps to preserve revenue from the realization of copyrights on these works. For instance, the YouTube Content ID system not only detects and blocks music plagiarism but also redirects the revenue generated from it to the original copyright holder. Methods for detecting music plagiarism which use artificial intelligence (specifically, neural network models) offer greater flexibility compared to classical methods. For example, plagiarism can be detected even when the musical instru ment or arrangement changes, as well as in the presence of noise. The drawback of these methods is the requirement of graphics processing units to run the neural networks. The article describes the differences between an audio recording, a musical audio recording, and a musical work. It out lines the problem of musical plagiarism detection and various methods for solving it. A comparison is made between classical methods that do not utilize neural network models and methods that incorporate them. The MelodySim neural network model, which accounts for melodic similarity, is examined alongside the dataset used for its training. The accuracy of its predictions at the segment level is investigated by creating an auxiliary dataset and opti mizing the binary classification decision threshold for segment pairs to increase the F1-Score metric value. The obtained results can be applied to improve both the predictions of MelodySim and the architecture of the model itself. An approach similar to the one discussed can be used to improve the prediction accuracy of other models with a similar architecture. Future research could be aimed at improving the MelodySim model as well as expanding the dataset.

Problems in programming 2026; 3: 111-116