Detecting Plagiarism in Music Clips Using Audio Signature Matching

With millions of music clips released online, plagiarism detection is essential to safeguarding creativity and protecting artists. Our model aims to address two types of plagiarism: (i) accidental plagiarism, in which an artist unknowingly reproduces elements of existing works, and (ii) intentional plagiarism, where an existing song is altered or significant portion of the song is reproduced without crediting the original creator. To solve this issue our model performs spectral analysis to extract Mel-frequency Cepstral Coefficients (MFCCs) and Chroma features from MP3 clips in the MagnaTagATune (MTT) dataset. These features are processed and compressed into compact representations called signatures using a Convolutional Autoencoder (CAE). Each signature serves as a unique representation of a song and is stored in a database. When a user uploads a clip, its signature is generated and compared with existing signatures in the database using Euclidean distance. The system retrieves the top 10 most similar signatures and flags the clip as plagiarized if the similarity exceeds a predefined threshold. Our model achieved an accuracy of 91 % and a loss of 0.032. The solution provides a powerful tool backed by a novel approach in music plagiarism detection to empower users to proactively check their work for accidental plagiarism, preventing future legal disputes. Simultaneously, it enables creators to identify unauthorized use of their work and provides them with relevant metadata about the plagiarizers to quickly identify them and pursue legal action.