Similar Music Pair (SMP) Dataset

• The Similar Music Pair (SMP) dataset is a collection of musically aligned segment pairs designed to capture localized similarity for tasks like plagiarism detection and variation modeling.
• It integrates diverse variants such as real-world plagiarism cases, JAZZVAR original–variation pairs, and MelodySim’s algorithmically augmented examples, each with unique curation pipelines.
• Detailed annotations, similarity metrics, and benchmark protocols in SMP enable precise segment-level retrieval and evaluation, advancing music information retrieval research.

Searching arXiv for the cited SMP-related papers and closely related work. The term Similar Music Pair (SMP) dataset denotes a dataset design in which each datum is a musically related pair, typically aligned at the segment level and intended to expose localized similarity rather than only whole-song identity. In recent music information retrieval (MIR) literature, the label is used both for a specific public dataset of real-world plagiarism and remake cases and, more broadly, for SMP-style corpora such as JAZZVAR and MelodySim. Across these formulations, the common objective is to support tasks such as music plagiarism detection, segment-level retrieval, similarity modeling, and, in the case of JAZZVAR, generative variation modeling through Music Overpainting (Go et al., 29 Jan 2026, Go, 10 Sep 2025, Row et al., 2023, Lu et al., 27 May 2025).

1. Terminological scope and dataset variants

A common misconception is that SMP denotes a single, fixed benchmark. The literature instead uses the term in at least three closely related ways. First, the 2025 and 2026 plagiarism-detection papers introduce the Similar Music Pair dataset as a collection of real-world commercial recordings linked by documented plagiarism disputes or authorized remakes, with time-aligned annotations of similar passages. Second, the JAZZVAR paper can be read as an SMP corpus because it consists of manually matched Original–Variation pairs extracted from jazz-standard lead sheets and solo piano performances. Third, the MelodySim work presents a melody-aware pair dataset constructed from Slakh2100 by generating melody-preserving audio variations and aligning segment pairs between originals and augmented versions (Go, 10 Sep 2025, Go et al., 29 Jan 2026, Row et al., 2023, Lu et al., 27 May 2025).

The distinction matters because the three resources target different forms of similarity. JAZZVAR emphasizes variation within performance practice: each pair links a 4-bar lead-sheet segment to a corresponding 4-bar performed variation. The plagiarism-focused SMP benchmark emphasizes real-world borrowing, including “Plagiarism Case” and “Remake” relations, and attaches segment timestamps to commercial recordings. MelodySim emphasizes melodic preservation under controlled augmentation, producing positive pairs that preserve melody while altering other musical factors. This suggests that SMP is best understood as a dataset paradigm centered on paired, aligned musical relatedness, rather than as a single data release (Row et al., 2023, Go, 10 Sep 2025, Lu et al., 27 May 2025).

2. Pair semantics and annotation model

JAZZVAR provides the most explicit formalization of an SMP corpus. It defines a dataset

$D=\{(O_i,V_i)\}_{i=1}^N,$

where $O_i\in\mathcal{O}$ is a 4-bar “Original” segment containing melody and chords from a jazz-standard lead sheet, and $V_i\in\mathcal{V}$ is a corresponding 4-bar “Variation” segment extracted from a solo piano performance of the same standard. Pair admission can be expressed through a similarity function

$S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$

with a manual threshold $\tau$ such that only pairs with $S(O_i,V_i)\ge\tau$ are kept. In practice, similarity was assessed by melodic alignment and harmonic outline similarity, and melodic deviation was formalized as

$$\text{AvgDev}(O,V)=\frac{1}{n}\sum_{j=1}^{n}(PC_j+D_j),$$

where $PC_j$ is the pitch-class difference and $D_j$ is the duration difference at the $j$-th aligned note. Pairs were chosen so that $O_i\in\mathcal{O}$0, with the threshold manually tuned (Row et al., 2023).

The plagiarism-oriented SMP dataset defines pair structure differently. It provides ground-truth pairs $O_i\in\mathcal{O}$1 in which $O_i\in\mathcal{O}$2 is known, by musicologists or legal findings, to share some musical content with $O_i\in\mathcal{O}$3. Each pair is accompanied by time-stamp annotations for similar segments in both recordings, and the 2026 formulation adds an acoustic index that assigns repeated instances of the same pattern to the same group. Annotators marked boundaries at the nearest downbeat and also recorded the musical element responsible—melody, harmony, rhythm, or vocals—although this element label is held in reserve for diagnostic analysis in SMP. The 2025 description reports start and end timestamps reviewed by at least two experts, but does not report a formal inter-annotator-agreement study (Go et al., 29 Jan 2026, Go, 10 Sep 2025).

MelodySim adopts a third annotation model. It creates positive pairs by taking time-aligned 10 s segments between an original audio rendering and one of three augmented versions derived from the same MIDI source, and negative pairs from inter-song segment combinations with no shared origin. During triplet construction, the anchor is a segment from one version, the positive is the same time window from another version, and the negative is a random segment from a different song. The pair relation is therefore algorithmically defined, but a listening study is used to validate that the positives do in fact preserve melody while changing other musical material (Lu et al., 27 May 2025).

3. Curation pipelines

JAZZVAR was built through a manual curation workflow centered on jazz standards. The authors gathered 234 MIDI/MusicXML lead sheets in 4/4 time from MuseScore, cleaned them to keep only the “head” section, collapsed repeats, and retained 22 standards for which matching performances existed. They manually searched Spotify and YouTube for solo-piano recordings, downloaded 760 audio files from 148 albums by 101 pianists, and obtained usable matching variations from 47 of those recordings. Audio was converted to MIDI using the high-resolution piano transcription model of Kong et al. (2022). Four-bar Originals were extracted from lead sheets, a Python GUI was developed for browsing and aligning candidate passages, and the AMT MIDI performances were manually scanned for passages with high melodic and harmonic resemblance, yielding 502 pairs in total. Quality control consisted of manual spot checks on AMT artifacts and manual correction of clearly mis-transcribed notes when necessary (Row et al., 2023).

The real-world SMP dataset was curated from documented plagiarism and remake cases. The 2025 paper describes real-world commercial recordings spanning multiple decades and genres, including officially recognized plagiarism disputes and artist-authorized remakes. Pair selection required documented melodic, harmonic, or rhythmic overlap significant enough to have led to legal disputes or explicit acknowledgement by artists. Expert annotators, described as musicologists or trained MIR engineers, marked the start and end timestamps of every similar segment in both tracks, with all annotations reviewed by at least two experts before inclusion. The 2026 paper further specifies a human-in-the-loop procedure in which initial downbeat times and a 4-bar segmentation grid were generated automatically by a beat-tracker and bar-counting algorithm, after which trained musicologists refined boundaries and verified each match (Go, 10 Sep 2025, Go et al., 29 Jan 2026).

MelodySim was constructed by augmentation rather than by mining commercial borrowing cases. It starts from 1,568 multi-track MIDI files from Slakh2100. A binary gradient-boosting classifier with 97 % accuracy on a held-out CMU dataset identifies the melody track using per-track features such as pitch range, polyphony rate, and note density, together with cross-track reference statistics. For each nonmelody track, a probability $O_i\in\mathcal{O}$4 governs note splitting, chord inversion, and chord arpeggiation. Additional operations include instrument re-assignment and track dropout, with melody, bass, and certain instrument classes protected in specified ways. After rendering with the Musyng soundfont, pitch shift, time shift, and tempo stretch are applied, and each audio file is cut into non-overlapping 10 s excerpts (Lu et al., 27 May 2025).

4. Statistical profile and representations

JAZZVAR contains 502 pairs, covering 22 distinct tunes, 47 distinct solo-piano recordings, and 35 pianists. Both Originals and Variations are 4 bars long, although performance tempo varies, and the tempo range is reported as roughly 100–200 BPM. The majority of segments remain in keys matching standard lead sheets, but some performer transpositions occur. The feature summary reported for Originals versus Variations is musically informative: Pitch-Class Entropy changes from $O_i\in\mathcal{O}$5 to $O_i\in\mathcal{O}$6; Pitch Range from $O_i\in\mathcal{O}$7 to $O_i\in\mathcal{O}$8; Mean Polyphony from $O_i\in\mathcal{O}$9 to $V_i\in\mathcal{V}$0; Number of Pitches per 4-bar segment from $V_i\in\mathcal{V}$1 to $V_i\in\mathcal{V}$2; and Pitch-in-Scale Ratio from $V_i\in\mathcal{V}$3 to $V_i\in\mathcal{V}$4. Originals are stored as manually edited single-track MIDI containing melody and chords, whereas Variations are automatically transcribed two-hand piano MIDI. Per-pair metadata include pair_id, standard_title, composer, lead_sheet_source, performance_artist, performance_album, performance_year, AMT_model, tempo_bpm, meter, key_signature, and bars (Row et al., 2023).

The real-world SMP benchmark is described with slightly different counts across publications. The 2025 paper reports 70 pairs, approximately 140 unique tracks, and a CSV or JSON table whose rows include Original Title, Comparison Title, Relation, Original Time, Comparison Time, and Pair #. The 2026 paper reports 72 original–comparison song pairs, variable numbers of segments per pair, and on the order of 1,200–1,500 ground-truth pairs, with exact total withheld pending public release. It also states that no genre accounts for more than 15 % of the pairs and that annotated segments are typically 1–4 bars, even though indexable segments in experiments are fixed to 4-bar windows. This suggests that the dataset definition or release state changed between the two publications, while preserving the core design of real-world, segment-annotated musical borrowing pairs (Go, 10 Sep 2025, Go et al., 29 Jan 2026).

MelodySim reports 1,568 originals, 3 augmented variants per original, and therefore 6,272 total audio files. The train/test split is 95 % / 5 %, corresponding to approximately 1,490 originals for training and 78 originals for testing. The test set contains 546 positive pairs and 546 randomly sampled inter-song negative pairs. Data are stored as PCM audio, such as 16 kHz mono WAV, with pre-extracted MERT embeddings also provided. On HuggingFace, the audio is organized as Track<ID>/version_k/segment_j.wav (Lu et al., 27 May 2025).

5. Similarity modeling and benchmark protocols

The 2025 SMP plagiarism paper defines a multi-feature similarity score over musically meaningful segments. Its feature set comprises Pattern Similarity $V_i\in\mathcal{V}$5 based on chromagram intersection, Musical Complexity $V_i\in\mathcal{V}$6 as the count of unique pitches, Rhythmic Correlation $V_i\in\mathcal{V}$7 as Jaccard similarity over quantized onset times, BPM Difference Ratio $V_i\in\mathcal{V}$8, and Chord Similarity

$V_i\in\mathcal{V}$9

where $S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$0 is Roman numeral similarity and $S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$1 is chord quality similarity. The full score is

$S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$2

Using this framework, the paper reports Segment Precision@100 = 98%, described as 98 correct retrievals out of 100, and song-level Top-1 Accuracy = 37.86% with Top-5 Accuracy = 62.86% on SMP (Go, 10 Sep 2025).

The 2026 MPD formulation broadens the evaluation regime. It defines symbolic pianoroll similarity

$S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$3

chord-sequence similarity via normalized edit distance,

$S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$4

rhythmic similarity

$S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$5

and an aggregate music-domain score

$S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$6

For audio features, it specifies a DTW distance

$S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$7

Its segment-level metric is Rec. 1s@k, which requires both top-$S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$8 retrieval and a start time within $S:\mathcal{O}\times\mathcal{V}\to\mathbb{R},$9 s of the annotation, with the same acoustic index. On SMP Timestamps, MERT achieves 25.6 / 43.0 / 51.0 for Rec. 1s@{1,5,10}; on Full Indices, MERT drops to 2.2 / 7.0 / 7.9, while the “Music” baseline reaches 4.0 / 12.3 / 15.4. For music-level retrieval on Covers80, the “Music” baseline reports mAP = 0.475 and MR1 = 13.46, compared with mAP = 0.150 and MR1 = 33.3 for MERT (Go et al., 29 Jan 2026).

JAZZVAR and MelodySim attach these pair structures to different learning problems. JAZZVAR introduces Music Overpainting, in which, given an Original segment $\tau$0, a system generates a variation $\tau$1 that preserves core melody and harmony while rearranging ornamentation, voicing, and rhythm. Its baseline is a Music Transformer with relative self-attention trained on concatenated Original and Variation token sequences, with cross-entropy loss on Variation tokens and a 90 % / 10 % train/validation split of 452 / 50 pairs. The paper reports a validation token-level perplexity of approximately 4.2 and notes that generated variations capture increased polyphony and pitch range but imperfectly preserve melodic contour. MelodySim, by contrast, uses a frozen MERT-v1-95M encoder, selected hidden layers $\tau$2, temporal down-sampling, a 1D-conv ResNet adaptation network, a triplet head with Euclidean distance, and a binary classifier. On its balanced test set, it reports Accuracy = 0.97; for the Similar class, P = 0.94, R = 1.00, F1 = 0.97; and for the Different class, P = 1.00, R = 0.94, F1 = 0.97 (Row et al., 2023, Lu et al., 27 May 2025).

6. Applications, access, and outstanding issues

The principal application of the real-world SMP dataset is Music Plagiarism Detection (MPD). The 2026 task definition states that MPD requires three capabilities: retrieving a plagiarized track from a very large library, localizing the exact time-aligned segments, and ideally identifying which musical element was copied. This explicitly differentiates SMP-based MPD from traditional cover-song identification and from audio fingerprinting, both of which operate at different granularities or with different assumptions. The same dataset also supports segment-level retrieval and ranking-based evaluation, while the 2025 paper additionally positions it for more general music similarity modeling, including partial matching and fine-grained similarity analysis within and across tracks (Go et al., 29 Jan 2026, Go, 10 Sep 2025).

JAZZVAR extends the SMP idea into generative and performance-oriented MIR. Beyond Music Overpainting, the paper lists expressive performance analysis and performer identification as potential applications, since each variation is traceable to a pianist. It also notes possible use in style transfer within jazz sub-styles, such as bebop versus modal. MelodySim focuses more narrowly on melodic similarity modeling for plagiarism detection, but its listening study with 12 participants, each rating 12 pairs on a 7-point Likert scale, supplies an empirical validation of the central assumption: positive pairs preserve melody while significantly altering other elements. The reported means are 4.23 ± 0.80 for overall similarity, 4.53 ± 0.84 for melodic similarity, and 3.94 ± 0.53 for non-melodic similarity in positive pairs, versus 2.00 ± 0.68, 1.90 ± 0.90, and 2.27 ± 0.22 for negative pairs (Row et al., 2023, Lu et al., 27 May 2025).

Access conditions differ substantially across SMP-style corpora. JAZZVAR is not publicly downloadable because of underlying audio and lead-sheet copyrights; access is by request to the authors and is limited to non-commercial research under “fair use” provisions, although a GitHub repository is intended to release the GUI, data loaders, preprocessing scripts, and Transformer code. The plagiarism SMP metadata is publicly available as a GitHub repository, but no audio files are redistributed, so researchers must align timestamps against their own licensed copies of commercial recordings. MelodySim is released on HuggingFace under CC-BY with no commercial restrictions, and includes both audio and pre-computed MERT features (Row et al., 2023, Go, 10 Sep 2025, Lu et al., 27 May 2025).

Several open directions recur across the literature. JAZZVAR proposes semi-automatic pair matching, extension to other genres such as classical and pop, use of Original–Variation pairs for audio–MIDI alignment benchmarks, learning joint embeddings of $\tau$3 for retrieval or style-similarity ranking, and unsupervised methods for discovering variation segments automatically. The plagiarism-oriented SMP work leaves open the challenge of scaling segment-level retrieval to much larger libraries without the sharp degradation observed in current learned models. Taken together, these works position the SMP dataset concept as a rigorous substrate for studying localized musical relatedness across symbolic, audio, and multimodal MIR settings (Row et al., 2023, Go et al., 29 Jan 2026).