Summary

MelodySim is a melody-aware music similarity model and dataset designed for plagiarism detection in music. It uses a novel melodic similarity approach based on a segment-wise detection model to identify similar melodies between music pieces.

Dataset Construction

The MelodySim dataset is created by augmenting the existing Slakh2100 MIDI dataset to generate melody-preserving variations. Augmentations include note splitting, arpeggiation, minor track dropout (except bass), and re-instrumentation, resulting in 1,710 synthesized pieces and 419 hours of music with four different versions each.

Model and Method

MelodySim employs a segment-wise melodic similarity detection model using a MERT encoder combined with a triplet neural network to capture melodic similarities accurately. The output is a decision matrix indicating potential plagiarism segments between two audio files.

Installation and Use

• Clone the GitHub repository and install dependencies via pip.
• Precompute MERT features from audio datasets.
• Train the model using provided scripts with specified dataset directories.
• Run inference to analyze similarity between two audio files and decide if they are similar or potentially plagiarized.

Performance

Testing on the MelodySim dataset reports high performance with an average accuracy of 97%, precision and recall both at 97%, demonstrating effective melodic similarity detection.

Intended Users

This tool is aimed at researchers and developers interested in music plagiarism detection, similarity analysis, and automated music information retrieval.