SaiPavankumar22/bird-sounds

birdcall-permissive — openly licensed bird audio (18,219 clips, 18.3 GB)

The training audio behind birdcall-v4, an offline 93-class bird call classifier. Every clip here is CC0, CC BY 4.0, or public domain — no NonCommercial, no NoDerivatives, nothing that would make the derived model unpublishable.

This is the filtered corpus: 18,223 permissive rows taken from an 18,487-row manifest. The 264 Xeno-canto British birdsong clips (CC BY-NC-ND / BY-NC-SA) were excluded on purpose and are not in this repository.

Contents

source	files	size	licence	in repo
iNaturalist	11,605 rows / 11,601 files	12.34 GB	CC0	yes
DCASE 2018 Task B	6,600	5.84 GB	CC BY 4.0	yes
NPS Rocky Mountain	18	0.13 GB	public domain	yes
British birdsong	264	—	CC BY-NC-ND / BY-NC-SA	excluded

Total: 18,219 unique audio files, 18.30 GB.

Layout

audio/
  inat/<binomial>/<sound_id>.<ext>          91 species, CC0
  dcase2018/dev/<set>/<clip_id>.wav         3 sets, CC BY 4.0
  nps_rocky_mountain/audio/*.mp3            public domain
manifests/
  permissive_only.csv    <- the 18,223 rows whose audio is here
  all_clips.csv          <- the full 18,487-row manifest incl. the 264 excluded
  inat_clips.csv         <- iNaturalist provenance: url, attribution, observer

Widest folder holds 2,200 files (the Hub's per-folder cap is 10,000), and each species lives in its own directory, so nothing needs repartitioning.

Provenance

Every row of manifests/permissive_only.csv carries:

column	what it is
clip_id	stable id (inat_<sound_id> for iNaturalist)
path	path on the training host
class_name / binomial	the species label
source	inaturalist / dcase2018 / nps_rocky_mountain / british_birdsong
license	the authoritative per-file licence
attribution	the credit string to reuse
observer	recordist — used for the recordist-disjoint evaluation split
url	source observation / dataset page

license and attribution are populated on 100 % of rows (verified: 11,605/11,605 iNaturalist, 6,600/6,600 DCASE, 18/18 NPS, 264/264 British).

Licence

The repository-level license: is CC BY 4.0, the most restrictive of the three permissive licences present — so honouring it satisfies every file. Per-file licences in manifests/ are authoritative and more precise:

• iNaturalist audio is CC0 (public domain dedication). Check the individual observation page before reuse; the licence column records what was current at download time.
• DCASE 2018 Task B is CC BY 4.0 — credit the DCASE challenge and indicate changes.
• NPS recordings are US federal public domain.

Not here: the 264 British clips. Their metadata appears in all_clips.csv only so that the exclusion is auditable.

Required attribution

Redistribute or reuse this audio only with the credit each licence demands. The attribution column in manifests/permissive_only.csv holds the per-clip string; the repo-level obligations are:

DCASE 2018 Task B — CC BY 4.0 (6,600 clips). You must give appropriate credit, provide a link to the licence, and indicate if changes were made. Cite the DCASE 2018 Task B page, which carries the citation for each constituent corpus (ff1010bird, warblrb10k, BirdVox-DCASE-20k):

Mesaros, A., Heittola, T., Dikmen, O., Virtanen, T. Sound event detection in real environments with application to bird audio detection. DCASE 2018 Workshop. https://dcase.community/challenge2018/task-bird-audio-detection

Suggested credit line:

Audio from the DCASE 2018 Task B bird-audio-detection dataset, licensed CC BY 4.0. Changes: filtered to permissive licences, re-encoded for storage.

iNaturalist — CC0 (11,605 clips). CC0 requires no attribution, but each clip's origin is recorded and linking it back is appreciated:

Audio from iNaturalist observations, dedicated to the public domain under CC0. https://www.inaturalist.org · per-clip observation URL in manifests/inat_clips.csv.

NPS Rocky Mountain — public domain (18 clips). No credit required; source: https://www.nps.gov/subjects/sound/soundlibrary.htm

References

• iNaturalist — open audio observations under CC0. https://www.inaturalist.org
• Creative Commons Zero v1.0 Universal. https://creativecommons.org/publicdomain/zero/1.0/
• DCASE 2018 Task B, bird-audio-detection — dataset page and required citations. https://dcase.community/challenge2018/task-bird-audio-detection
• Creative Commons Attribution 4.0 International. https://creativecommons.org/licenses/by/4.0/
• National Park Service, Rocky Mountain Sound Library (US federal public domain). https://www.nps.gov/subjects/sound/soundlibrary.htm
• BirdVox — acoustic sensor data underlying part of the DCASE corpus. https://www.birdvox.org
• Xeno-canto — bird recordings; the British subset used by the research-only model versions is CC BY-NC-ND / BY-NC-SA. https://xeno-canto.org
• British birdsong dataset (Kaggle), the local copy of that subset. https://www.kaggle.com/datasets/rtatman/british-birdsong-dataset

Creative Commons NonCommercial and NoDerivatives terms are the reason those 264 recordings are excluded here: https://creativecommons.org/licenses/by-nc-nd/3.0/ · https://creativecommons.org/licenses/by-nc-sa/3.0/

Loading

from pathlib import Path
import soundfile as sf

for p in Path("audio/inat").rglob("*.wav"):
    y, sr = sf.read(p)
    break

Or with datasets:

from datasets import load_dataset
ds = load_dataset("<your-hf-username>/birdcall-permissive", split="train",
                  streaming=True)

Relationship to the model

This audio produced birdcall-v4: frozen Perch 2.0 backbone (Apache-2.0) plus a linear head trained on 16,591 of these clips, reaching 0.846 species top-1 under 3-fold clip-grouped cross-validation and 0.833 recordist-disjoint.

1,567 clips here failed to decode (known libmpg123 junk-header failures) and 65 are a frozen test set excluded from training — which is why 18,223 rows become 16,591 training clips.

Reproducing the pull

python tools/inat_download.py \
    --out-dir ~/bird_dataset/raw/inat \
    --manifest ~/bird_dataset/manifests/inat_clips.csv \
    --max-per-species 100000 --max-pages 40 --per-observer 0
python tools/build_manifest.py --dataset-root ~/bird_dataset \
    --out manifests/all_clips.csv      # prints the licence audit

The iNaturalist pool for these species is exhausted: a full 93-species dry run with those flags reports 0 new.

Derived model

The weights trained on this audio are published as SaiPavankumar22/Bird-finder, which ships all four model versions side by side plus the evaluation that compares them. Any model built from this data inherits the obligations of the license column of the clip it was trained on — see Licence above.

Disclaimer

Not a general bird-identification service — 91 species is a small slice of the birds that exist. Confidence flags report model agreement, not correctness.