Recurring MIDI phrases with source coordinates and separate drum patterns.
SaMuGeD began as a MIDI sampling project: find repeated musical ideas, collect them as reusable phrases and explore them in music production. Earworms carries that idea into a dataset and a loop player.
flowchart LR
S[Source MIDI] --> M[Melodic parts]
S --> D[Drum parts]
M --> R[Find and verify repeats]
D --> R
R --> P[Phrase MIDI and source coordinates]
classDef data fill:#e6e8f5,stroke:#6b6e91,color:#222437
classDef action fill:#e2eeea,stroke:#567e72,color:#23342e
class S,P data
class M,D,R action
The primary aligned_closed release has 95,077 phrases, 50,566 melodic and 44,511 percussion, with complete artifact audits and closed selection replay on a declared 256 source sample. The fixed window baseline of the research note (94,950 phrases, replayed on all 16,995 successful sources) is kept on the hub only as the reference archive and is no longer a dataset configuration.
Local corpus expansion tooling adds indexed source metadata and rights evidence, including separate corpus and score declarations, MIDI copyright notices and unknown composition rights. The expansion remains local and is separate from the published release above.
flowchart LR
P[Verified phrases] --> A[Parquet and MIDI archives]
P --> T[Top 50 analytics]
T --> L[Source derived audio loops]
F[Familiar song evidence] --> L
L --> W[Continuous playback and DAW downloads]
classDef data fill:#e6e8f5,stroke:#6b6e91,color:#222437
classDef action fill:#e2eeea,stroke:#567e72,color:#23342e
classDef evidence fill:#f2ebe0,stroke:#968261,color:#3b3224
class P,A,T data
class L,W action
class F evidence
The Space includes top 50 listening collections, ten familiar song selections and Tool’s ostinatos with 69 riffs and 36 drum patterns from 28 songs. Source drums play by default where available. Tool offers 59 aligned three mode comparisons. Download 48 kHz stereo WAV, FLAC or loop MIDI. The separate atlas provides recurrence counts and note plots. Its audio uses the same SoundFont renderer. Downloaded Tool arrangements are interface supplements, outside the dataset.
Start with the interactive Schism example to see repeated notes become a loop. All listening audio uses ColomboGMGS2 17.02 Vanilla through FluidSynth.
Rendering notes include a reproducible six phrase comparison of complete instrument banks and MIDI engines.
There are no listener labels for these fragments. Repetition does not prove catchiness or recognition. Familiar song evidence applies to songs, not these exact MIDI fragments. Filename identities, search limits and duplicate arrangements affect coverage.
Corpus expansion and searchable metadata describes source conditions, bounded imports and the new catalog.
uv venv .venv --python 3.12
source .venv/bin/activate
uv pip install -r requirements-research.lock -r requirements-publication.lock
pytest -q
python -m samuged.cli build --source 'datasets/Lakh MIDI Clean' \
--output research_local/my_run --algorithm aligned_closed \
--workers 4 --percussion --recover-invalid-keysUse a new output directory. Add --limit 128 for a pilot. Source MIDI is not bundled.
Methods and verification · Publication and rendering · Release status · Consumer guide · Familiar song selection
The original roadmap connects music structure analysis, phrase extraction, feature search and DAW use. It records the project's starting ideas, not a list of completed features.
| Original direction | Where it stands |
|---|---|
| Find repeated phrases | The current dataset exports melodic and drum patterns with verified source occurrences. |
| Explore segmentation and phrase boundaries | SF segmenter experiments are retained. The current method uses its own candidate and boundary rules. |
| Group and search musical patterns | SimilarMidis retains the earlier feature similarity application and clustering notebook. |
| Reuse phrases in music production | The player offers looping audio and MIDI downloads for DAWs. A dedicated DAW plugin is not part of this release. |
The original planning notes retain the earlier tasks on bar structure, note histograms, segment lengths and silence settings, alongside historical setup instructions.
| Author | Contact |
|---|---|
| Peiyi Wu | pewu10205@uit.no |
| Asle Fjæran Øren | asleoren@gmail.com |
| Shayan Dadman | shayan.dadman@uit.no |
| Almaz Ermilov | almaz.ermilov@uit.no |
Use Cite this repository in the GitHub sidebar, or copy the BibTeX below. Download BibTeX · Citation metadata.
@misc{wu2026samuged,
author = {Wu, Peiyi and Øren, Asle Fjæran and Dadman, Shayan and Ermilov, Almaz},
title = {{SaMuGeD} Earworms (Ostinato / Catchy musical hooks)},
year = {2026},
version = {0.3},
url = {https://huggingface.co/datasets/AlmazErmilov/samuged-recurring-phrases},
note = {Research dataset release}
}Code uses MIT. The Lakh and PDMX configurations follow their declared CC BY 4.0 collection licenses (Lakh, PDMX) and the MAESTRO configuration follows CC BY-NC-SA 4.0. Underlying composition and arrangement attribution is incomplete. No independent rights clearance is claimed. External evaluation datasets and original commercial recordings are excluded.