Birds / Global multi-site passive acoustic dataset

WABAD maps bird calls at scale; it does not map meaning

Journal paper and latest Zenodo archive version 2026 · Note reviewed 2026-10-23

WABAD is a promising large-scale bird-acoustic reference. The publisher reports 5,047 minutes of audio with 91,931 species-level annotated vocalizations from 1,192 species, 72 recording sites, 29 locations and 13 biomes. The latest Zenodo record lists 21.3 GB. This scale could help researchers test whether a sound model detects and identifies bird vocalizations across sites. The reviewed task description concerns vocalization occurrence and species labels; it does not test a receiver's response or what a call means.

What we reviewed · Primary publisher abstract and current Zenodo v4 metadata review; no archive acquisition or reproduction

Reviewed the publisher abstract and data-availability statement, plus the official current Zenodo v4 record, version history and file inventory. The full article body, supporting-information PDF, archive contents, metadata schema, labels, evaluation split and code were not inspected; no data or models were acquired or run.

Not yet independently reviewed by a bioacoustics or animal-communication specialist.

Study at a glance

Scale reported by publisher: 5,047 minutes of audio and 91,931 species-level annotated vocalizations.
Coverage reported by publisher: 1,192 bird species from 72 recording sites and 29 recording locations across 13 biomes.
Current archive record: Zenodo v4, published 2026-06-02; 21.3 GB listed across audio, annotations and metadata. No files downloaded or verified.
What this could test: Bird-vocalization detection/identification and, after split review, generalization across sites or regions.
What it does not test: Independent behavioral context, receiver response, comprehension or call meaning.
Rights status: Publisher abstract says dataset CC BY 4.0; current Zenodo v4 says CC BY-NC 4.0. Unresolved; metadata-only.

Reading location: Pérez-Granados et al. (2026), DOI 10.1002/ecy.70317; Zenodo DOI 10.5281/zenodo.20513304, v4, published 2026-06-02. The publisher data statement points to the earlier Zenodo DOI 10.5281/zenodo.17293588.

What this does — and doesn’t — establish

The journal abstract states that the dataset is CC BY 4.0, but the current Zenodo v4 record states CC BY-NC 4.0. The archive's version history and publisher data statement point to different versions; the reuse discrepancy is unresolved. Source-recording provenance and per-file conditions were not audited. We did not inspect the underlying audio, annotations, metadata, label quality, grouping keys or any train/test split. WABAD therefore remains a metadata-only research candidate, not an admitted training set or a Talk2Nature benchmark.

What we can build on

A large, multi-site bird dataset can teach us how well software finds and sorts vocalizations, and whether that skill travels across places. It cannot by itself teach us what a bird is communicating. Keep detected sound, human-observed context, an animal's measured response and any model prediction in separate records. Start with a small, rights-cleared detection question before attempting a meaning or conversation claim.

A practical starting point

Metadata-only candidate for bird-vocalization detection; not admitted to Talk2Nature training or a communication benchmark.

Available starting point: Publisher article abstract and official Zenodo v4 record 20513304. No paper body, archive, audio, annotations, code or model was acquired.

Reuse limits: The publisher abstract reports CC BY 4.0 for the dataset, while current Zenodo v4 reports CC BY-NC 4.0. Resolve this discrepancy and review source-record/per-file conditions before acquisition. Commercial permission is not cleared.

Our proposed experiment: No data were acquired or analyzed, no model was trained and no result was reproduced. A future detector test requires a rights-cleared version, inspected grouping and labels, an independent reviewer, a simple baseline and a locked site-held-out evaluation. Detection performance would not establish animal behavior, comprehension or meaning.

The next reading task

Resolve the article/archive license mismatch against the exact versions and obtain an authoritative reuse statement; inspect the supporting metadata and schema, label definitions, site/device/session grouping and evaluation split; then ask an independent bioacoustics reviewer whether a bounded bird-detection test would answer a useful question. If those gates pass, predefine a simple baseline and site-held-out evaluation before any model run. Do not use WABAD to claim call meaning or commercial readiness.

Go to the sources

WABAD data paper, Ecology (2026) ↗
Pérez-Granados et al., WABAD: A World Annotated Bird Acoustic Dataset for Passive Acoustic Monitoring, Ecology 107(2), e70317. DOI 10.1002/ecy.70317.
WABAD latest Zenodo dataset record, v4 ↗
Official current version of the World Annotated Bird Acoustic Dataset, with 21.3 GB of audio and associated metadata/annotations listed on the record.

Original editorial note by Talk2Nature contributors. AI-assisted, source-grounded synthesis; not a peer review or an independent replication. Linked works retain their own rights.