Download README.md from ben2002chou/notesep-benchmarks: direct link, hf CLI and curl.
- Browser
- Download file 2.03 kB
-
https://huggingface.co/datasets/ben2002chou/notesep-benchmarks/resolve/main/README.md
- Command line
-
hf download hf://datasets/ben2002chou/notesep-benchmarks/README.md
-
curl -L -o README.md https://huggingface.co/datasets/ben2002chou/notesep-benchmarks/resolve/main/README.md
pretty_name: NoteSep separation and editing benchmark
license: other
license_name: notesep-benchmark-source-licenses
license_link: >-
https://huggingface.co/datasets/ben2002chou/notesep-benchmarks/blob/main/SOURCES.md
tags:
- audio
- music
- source-separation
NoteSep benchmark
144 short recordings with aligned scores and 3,473 isolated note references, for evaluating separation and note editing. Download size is about 152 MB.
Both tasks share the same audio:
| File or folder | Contents |
|---|---|
evaluation.json |
Note requests, crop settings and scoring definitions. |
editing.json |
Fixed gain, pitch, timing and tuning edits, with their units. |
pieces/ |
Mixtures, aligned scores, isolated-note archives and timing metadata. |
manifest.json |
Download inventory for notesep download-data. |
SOURCES.md, score-credits.json, CC0.txt |
Instrument and score credits, source terms and transformations. |
mutopia_*.mid |
Original MIDI for the two Mutopia score editions used here. |
From the NoteSep repository:
uv run notesep download-data --repo-id ben2002chou/notesep-benchmarks --output data/benchmark
For separation, use a piece's mixture.wav and score.json. For evaluation, use the measured acoustic endpoints and per-request crops in evaluation.json; the convenience score contains MIDI note-offs.
For editing, use the same isolated notes and the requests in editing.json. Gain is in dB, pitch in semitones, timing in milliseconds, and tuning in cents. Construct edited mixtures by summing the edited note with the other notes. These files provide the original clean notes and edit specifications, rather than a separate collection of precomputed edited mixtures.
Keep this benchmark out of training. Audio is stored at 24 kHz; NoteSep inference resamples it to 16 kHz.
This release uses Swagbass for bass and direct Mutopia editions for the Bach BWV 871 and Beethoven Op. 13 excerpts. See SOURCES.md for credits and edition details.