BAM_subset_demo / README.md
sammlapp's picture
Upload folder using huggingface_hub
07fede2 verified
|
Raw History Blame Contribute Delete
3.9 kB
metadata
license: cc-by-4.0
task_categories:
  - audio-classification
tags:
  - bioacoustics
  - birds
  - passive-acoustic-monitoring
  - multi-label
  - education
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: test
        path: data/test-*.parquet

BAM 3-species bird sound dataset (Cameroon)

A small multi-label bird sound classification dataset for teaching. It is a subset of the BAM site (Bamenda Highlands, Big Babanki, Cameroon) from WABAD, the World Annotated Bird Acoustic Dataset for Passive Acoustic Monitoring.

Each example is a 3-second, 24 kHz mono clip (FLAC) from a passive recorder in tropical montane forest. The task is to detect which of three target species are vocalizing.

Classes

Label Common name Scientific name
0 Orange-tufted Sunbird Cinnyris bouvieri
1 African Hill Babbler Sylvia abyssinica
2 Great Blue Turaco Corythaeola cristata

labels is a (possibly empty) list of class indices. A clip can contain several target species. An empty list means none of the three targets is present (a negative clip).

Splits

clips with a target negative of which have other species
train 866 425 441 200
test 472 95 377 303

Clips per class (a clip can have more than one):

Orange-tufted Sunbird African Hill Babbler Great Blue Turaco
train 259 140 39
test 62 25 9

Train is not a random sample of the original split. It contains every original train clip with a target species, every original train clip with no annotated vocalization (241), and 200 randomly chosen (seed 0) train clips that contain only other species. Most other train clips were dropped, so the train set has far fewer "other species" negatives than the full recording would.

Test is the complete BAM test split (472 clips) of the clip-level train/test metadata this subset was built from. Most test clips are negatives for these three classes, and many contain other species (Yellow-breasted Boubou, Yellow Bishop, apalis, other sunbirds, etc.). Test negatives are therefore harder and more numerous than train negatives, and the class balance differs between splits.

Fields

  • audio: 24 kHz mono audio (resampled from 48 kHz, stored as 16-bit FLAC)
  • filename: original clip file name, BAM_<YYYYMMDD>_<HHMMSS>_<clip index>.wav
  • labels: list of class indices (see above)
  • species_scientific: scientific names of the target species present
  • has_other_species: whether the clip also contains a non-target species that was annotated
  • recording_date, recording_time: parsed from the file name
  • clip_index: index of the 3 s clip within its source recording
from datasets import load_dataset
ds = load_dataset("<user>/<repo>")

Source and attribution

All audio and annotations come from WABAD:

Pérez-Granados, C., et al. WABAD: A World Annotated Bird Acoustic Dataset for Passive Acoustic Monitoring. Zenodo, v3, 2025. https://doi.org/10.5281/zenodo.17293588 (also published in Ecology, 2026).

The BAM recordings were contributed by the site contact listed in WABAD (tomasz.osiejuk@amu.edu.pl; Song Meter SM3 recorder, 2015). Please cite WABAD, and credit the original data providers, if you use this dataset.

Licensed under CC BY 4.0, the same license as the original dataset. Changes made: selected a subset of clips, resampled audio to 24 kHz, converted to FLAC, and reformatted the labels.

Limitations

Labels come from expert annotation of the original recordings and may miss faint vocalizations. Great Blue Turaco has few examples (39 train, 9 test). Intended for educational use, not as a benchmark.