Music Plagiarism Detection Eval

---
name: music-plagiarism-detection-eval
description: Evaluates a model's ability to detect plagiarized or remixed segments within audio tracks by computing segment-level musical similarity and attributing similarities to specific elements like melody, chords, and vocals. Use when the user wants to benchmark on Similar Music Pair, or asks about evaluating this task. Reports similarity score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2601.21260
  bibtex_key: [REDACTED]
  confidence: medium
---

# music-plagiarism-detection-eval

> Music Plagiarism Detection: Problem Formulation and a Segment-based Solution — Go et al. (2026) (arXiv:2601.21260, 2026)

## What this evaluates

Evaluates a model's ability to detect plagiarized or remixed segments within audio tracks by computing segment-level musical similarity and attributing similarities to specific elements like melody, chords, and vocals.

## Datasets

- **Similar Music Pair** — total ?; splits: (unstated); repo https://github.com/Mippia/ICASSP2026-MPD

## Metrics

- `similarity score` **(primary)** — range: [0, 1]
  - Computed independently for pianoroll, onset rhythm, and chord elements, then aggregated via a weighted combination to produce a final segment-level similarity score.

## Input / output format

**Input**: Raw audio segments or their transcribed representations (pianoroll, onset rhythm, chord, vocal tracks).

**Output**: Per-segment similarity scores for each musical element, a combined weighted similarity score, and element-level attribution labels (melody, chord, vocal).

## Scoring recipe

```python
def compute_similarity(seg1, seg2):
    # Extract transcriptions
    piano1, piano2 = get_pianoroll(seg1), get_pianoroll(seg2)
    rhythm1, rhythm2 = get_onset_rhythm(seg1), get_onset_rhythm(seg2)
    chord1, chord2 = get_chords(seg1), get_chords(seg2)
    
    # Compute element-wise similarities
    sim_piano = compute_pianoroll_similarity(piano1, piano2)
    sim_rhythm = compute_rhythm_similarity(rhythm1, rhythm2)
    sim_chord = compute_chord_similarity(chord1, chord2)
    
    # Weighted combination
    final_score = w1 * sim_piano + w2 * sim_rhythm + w3 * sim_chord
    return final_score
```

## Common pitfalls

- Acoustic features in audio-only models can mask plagiarism in specific musical components if performance differs between tracks.
- Algorithmic similarity methods may be outdated and lack robustness compared to deep learning approaches.
- Element-level attribution requires explicit transcription models rather than end-to-end audio encoders.

## Evidence (verbatim from paper)

> We compute pianoroll similarity, onset rhythm similarity, and chord similarity with segment information. Each element is calculated independently to derive the final similarity score through weighted combinations.

## Citation

```bibtex
@misc{go2026musicplagiarism,
  title={Music Plagiarism Detection: Problem Formulation and a Segment-based Solution},
  author={Go et al. (2026)},
  year={2026},
  note={arXiv:2601.21260}
}
```

- arXiv: 2601.21260