We propose MelodySim, a melody-aware music similarity model and dataset for plagiarism detection. First, we introduce a novel method to construct a dataset with focus on melodic similarity. By augmenting Slakh2100; an existing MIDI dataset, we generate variations of each piece while preserving the melody through modifications such as note splitting, arpeggiation, minor track dropout (excluding bass), and re-instrumentation. A user study confirms that positive pairs indeed contain similar melodies, with other musical tracks significantly changed. Second, we develop a segment-wise melodic-similarity detection model that uses a MERT encoder and applies a triplet neural network to capture melodic similarity. The resultant decision matrix highlights where plagiarism might occur. Our model achieves high accuracy on the MelodySim test set.
git clone https://github.com/AMAAI-Lab/MelodySim.git
cd MelodySim
pip install -r requirements.txtThe MelodySim dataset contains 1,710 valid synthesized pieces originated from Slakh2100 dataset, each containing 4 different versions (through various augmentation settings), with a total duration of 419 hours.
First, pre-compute the MERT features for the melodysim dataset:
python data/extract_mert.py -tracks-dir /path/to/melodysim/wav/dataset/train -out-dir /path/to/melodysim/mert/dataset/train
python data/extract_mert.py -tracks-dir /path/to/melodysim/wav/dataset/validation -out-dir /path/to/melodysim/mert/dataset/validation
Then, run the training script. The training results (checkpoints) will be saved in the results folder with the following command:
python train.py -dataset-root-dir /path/to/melodysim/mert/dataset -result-dir-basename results/training-run
You can run inference.py to run the model on two audio files, analyzing their similarity and reaching a decesion on whether or not they are the same song. We provide a positive pair and a negative pair as examples. Try out
python inference.py -audio-path1 ./data/example_wavs/Track01968_original.mp3 -audio-path2 ./data/example_wavs/Track01976_original.mp3 -ckpt-path path/to/checkpoint.ckpt
python inference.py -audio-path1 ./data/example_wavs/Track01976_original.mp3 -audio-path2 ./data/example_wavs/Track01976_version1.mp3 -ckpt-path path/to/checkpoint.ckpt
Feel free to play around the hyperparameters
-window-len-sec,-hop-len-sec(the way segmenting the input audios);--proportion-thres(how many similar segments should we consider the two pieces to be the same);--decision-thres(between 0 and 1, the smallest similarity value that we consider to be the same);--min-hits(for each window in piece1, the minimum number of similar windows in piece2 to assign that window to be plagiarized).
If you are interested in testing performance (precision, recall, F1), you can run the following commands.
python test.py -ckpt-path path/to/checkpoint.ckpt -tracks-dir /path/to/melodysim/wav/dataset/test --save-results-dir-basename results/testing --max-num-anchors 78
The argument --max-num-anchors specifies the number of testing anchors. In melodysim dataset, there are 4 versions for each anchor, which means that there are 6 positive pairs for each anchor. test.py considers all positive pairs plus the "anchor vs. anchor" case; for the negative pairs, it will randomly select the same amount as the positive pairs for testing.
Our testing results are as follows:
| Precision | Recall | F1 | |
|---|---|---|---|
| Different | 1.00 | 0.94 | 0.97 |
| Similar | 0.94 | 1.00 | 0.97 |
| Average | 0.97 | 0.97 | 0.97 |
| Accuracy | 0.97 |
If you find this work useful in your research, please cite:
@article{lu2025melodysim,
title={Text2midi-InferAlign: Improving Symbolic Music Generation with Inference-Time Alignment},
author={Tongyu Lu and Charlotta-Marlena Geist and Jan Melechovsky and Abhinaba Roy and Dorien Herremans},
year={2025},
journal={arXiv:2505.20979}
}