notesep-aso-data / README.md
ben2002chou's picture
Add complete dataset instructions
078a76c verified
|
Raw History Blame Contribute Delete
1.87 kB
---
pretty_name: NoteSep ASO training data
license: other
license_name: notesep-source-licenses
license_link: https://huggingface.co/datasets/ben2002chou/notesep-aso-data/blob/main/SOURCES.md
tags:
- music
- audio
- source-separation
---
# ASO training data
Source-complete mixtures and their isolated notes for training NoteSep's ASO.
Each crop contains every acoustically active source, including release tails.
- `synthetic/index.sqlite`: mixture membership, note timing and train/validation splits.
- `synthetic/shards/`: original float32 mixture and note banks, distributed as gzip files.
- `synthetic/AUTHORITY.json`: measured acoustic endpoint definitions.
- `synthetic/proposals.sqlite`: the score queries used to generate NoteGrab estimates.
- `recorded/`: recorded mixtures and aligned isolated-note targets.
- `manifest.json`: download inventory and automatic path relocation.
- `example/`: one complete crop as ordinary float32 WAVs and note labels.
Run `notesep download-data` to verify and restore the original files. Then use
`training/prepare_aso.py` to generate NoteGrab estimates locally before training.
The estimates are not included in this dataset. Training uses only the train
and validation selections; the original index also records its internal test
selection, which is not admitted to training or proposal generation.
Download size: about 71 GB. Restored audio and metadata occupy about 470 GB.
Allow roughly 550 GB of free disk space during setup. The example contains
11 aligned note waveforms whose sum reconstructs its mixture.
## Sources and licenses
See [SOURCES.md](SOURCES.md) for instrument, score and recording credits, transformations and source-specific terms. Original notices are included in `licenses/` and source information in `sources/`. Noncommercial and share-alike terms apply to the corresponding source material.