Birds of northeastern North America / Strongly annotated passive-acoustic dataset

A large forest-bird soundscape set can test detection, not conversation

Scientific Data Early Article and Zenodo dataset v2 2026 · Note reviewed 2026-10-23

The latest Zenodo v2 record describes 1,302 ten-minute forest soundscapes—about 217 hours—from 104 sites in four northeastern U.S. parks, with approximately 183,000 vocalization annotations. The version-pinned site_metadata.csv has 104 unique sites; recording_metadata.csv independently has the same 104 site IDs. Its site counts are ACAD 31, HBEF 24, KAWW 24 and MABI 25. The landing-page opening summary says KAWW has 20 sites, so the headline total is supported while that park breakdown remains inconsistent. The v2 README says 88 bird species, matching the publisher abstract and earlier v1; however, species_metadata.csv has 96 rows marked IsBird=Yes, all with nonzero annotation counts totaling 182,633. The species-count discrepancy therefore remains unresolved. This is a promising field-acoustics resource for event detection, localization and species classification, not a communication dataset.

What we reviewed · Publisher abstract and rights notice, complete Zenodo v1/v2 metadata, and checksum-verified aggregate inspection of the four-file v2 metadata sample; no audio or annotation rows reviewed

Reviewed the publisher title, abstract, publication notice and rights statement; the full article and supplement remain unreviewed. Read the official Zenodo v1/v2 record descriptions and complete v2 API manifest. Acquired only ReadMe.txt, site_metadata.csv, recording_metadata.csv and species_metadata.csv from v2. Verified provider MD5 declarations against bytes and recorded local SHA-256 values in an ignored receipt. The aggregate script emits no raw site IDs or coordinates. No recordings, annotation rows, code or models were acquired or run.

Not independently reviewed by an ornithologist, bioacoustics specialist or ML/statistical-evaluation reviewer.

Study at a glance

Latest release scale: Zenodo v2 reports 1,302 ten-minute recordings and approximately 183,000 annotations across four northeastern U.S. parks.
Site-count reconciliation: The checksum-verified site table and recording table each contain the same 104 unique sites. Site-table counts are ACAD 31, HBEF 24, KAWW 24 and MABI 25; the landing-page opening breakdown lists KAWW 20. The total is corroborated; the KAWW subcount remains inconsistent.
Species-count reconciliation: The v2 README and publisher abstract report 88 bird species. The species table contains 96 IsBird=Yes rows with nonzero annotation counts totaling 182,633. The difference remains unresolved.
Annotation target: Expert-reviewed time/frequency boxes with bird-species codes; not caller identity, behavior, addressee or receiver response.
Talk2Nature inspection: Only four small metadata files were acquired; their source MD5s and local SHA-256 receipt were verified. No audio or annotation rows were acquired. Coordinates and raw row-level identifiers were not exported.

Reading location: Scientific Data DOI 10.1038/s41597-026-08226-8; current Zenodo v2 record 20038954 / DOI 10.5281/zenodo.20038954; superseded v1 record 18041381.

What this does — and doesn’t — establish

The annotations identify vocalization events and species, not individual callers, observed social behavior, receiver response or call meaning. The site total of 104 is corroborated by two metadata tables and the README, but the landing page lists KAWW as 20 while site_metadata.csv has 24. The v2 README/publisher abstract say 88 bird species while the species table marks 96 rows as birds; the annotation total for those rows is 182,633, and the discrepancy is unresolved. The current record says annotators may combine consecutive same-species calls when pauses are under two seconds, and faint or unidentifiable events may remain unlabeled. Sampling differs by subset: SIMR is tied to point-count surveys across sites, while ACAD and MABI use selected clear mornings during peak breeding season. These choices matter for event recall, class balance and site/time holdouts. The archive is concentrated in four parks in one region and does not establish performance on household parrots, phones or other taxa. Site coordinates are present in the metadata; no values were published or mapped. The dataset record declares CC BY 4.0, while the article text is CC BY-NC-ND 4.0; neither statement is an independent audit of every component or source-media provenance.

What we can build on

A large, strongly annotated bird collection can help test whether software finds and classifies vocalizations across field recordings. It cannot show what a bird meant or whether another bird understood it. Keep this detection track separate from the Egyptian-fruit-bat context question, V33DA caller attribution and the provisional parrot observation workflow.

A practical starting point

Four-file, checksum-verified metadata sample inspected for aggregate counts only; not admitted to Talk2Nature training, species detection or communication research.

Available starting point: Publisher abstract and rights metadata; Zenodo v1/v2 record descriptions; and the v2 ReadMe.txt, site_metadata.csv, recording_metadata.csv and species_metadata.csv files acquired to ignored local storage. No audio, annotation rows, code or model files were acquired.

Reuse limits: The current Zenodo v2 record declares CC BY 4.0 for the dataset; the article text separately carries a publisher-stated CC BY-NC-ND 4.0 license. Verify exact version, source-media provenance and location conditions before use. Record-level terms are not an independent component audit.

Our proposed experiment: Aggregates reconcile the 104-site total but leave a KAWW park-count difference and 88-versus-96 species discrepancy. No model was run, no labels were validated and no benchmark result is claimed. Any future detection study needs expert review and site/park/year-separated groups with an untouched holdout.

The next reading task

Ask the authors to clarify the README/species-table definition difference and the KAWW landing-page/site-table difference. Before any audio or annotation acquisition, obtain independent bioacoustics and ML/statistical review of the minimum useful task, site/park/year grouping, class coverage, coordinate handling and a separate locked test. The four metadata files remain local and are not admitted to training or biological inference. Keep the full release metadata-only until a bounded rights/privacy plan and independent review pass.

Go to the sources

Panwar et al. Eastern North American forest-bird soundscape data descriptor ↗
Scientific Data Early Article, published 11 September 2026. Its abstract reports 217 hours and about 183,000 annotated vocalizations across 88 bird species. The publisher states CC BY-NC-ND 4.0 for the article text.
Northeastern forest-bird soundscapes, Zenodo v2 ↗
Current dataset record, Zenodo 20038954 / DOI 10.5281/zenodo.20038954. The record describes 1,302 ten-minute recordings from 104 sites in four parks and about 183,000 annotations across 96 bird species; it declares CC BY 4.0.
Northeastern forest-bird soundscapes, Zenodo v1 ↗
Superseded v1 dataset record, Zenodo 18041381. It reports approximately 176,000 vocalizations across 88 bird species; the current v2 record reports about 183,000 across 96 species.

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.