Birds / Passive acoustic dataset

CEB can benchmark bird sound detection, not what a call means

Dataset version publication year 2026 · Note reviewed 2026-10-23

CEB v1.1 is a large library of bird recordings with human-marked sound events: 20,826 audio files and 92,857 annotations across 256 eBird species codes and 21 vocalization types. It can help test whether software finds and sorts labeled calls in recordings. It contains no independent observations of a listener's response or ground truth about what a bird meant, so it cannot test translation. The field soundscape recordings are more relevant to passive-monitoring workflows than the separate focal Xeno-Canto recordings.

What we reviewed · Official Zenodo v1.1 record, version notes, schema and API file inventory reviewed; annotation rows, media and model remain uninspected

Reviewed the official Zenodo v1.1 landing page, published file inventory, component-specific rights statement and schema.json browser preview. The v1.1 notes identify annotation and taxonomy corrections; the prior v1 record 20762099 is superseded. Re-read the official Zenodo API inventory on October 23, 2026: train_soundscape.csv is 8,559,233 bytes (provider MD5 55b5e3f99566c5bb3ba4fb5b833424da) and test_soundscape.csv is 2,401,049 bytes (provider MD5 f572155ae41714e17b218e95d97b1ce9). These checksums are upstream metadata, not locally verified. No CSV rows, audio, code or weights were downloaded, played, parsed or executed; no detector was evaluated.

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

Study at a glance

Scale: The official v1.1 record lists 20,826 audio files, 92,857 event annotations, 256 eBird species codes and 21 vocalization types. The packaged collection is 49.7 GB.
Soundscape subsets: train_soundscape has 18,469 field recordings and 62,298 weak annotations that mark an approximate time point; test_soundscape has 147 field recordings and 15,064 stronger labels with event boundaries.
Focal recordings: train_xenocanto has 2,210 foreground recordings and 15,495 annotations, re-labeled with event boundaries and vocalization types. Each recording retains its own upstream license.
Label meaning: Labels include sound/vocalization categories such as song, contact call, alarm call and interaction call. These dataset labels are not independent tests of receiver comprehension or semantic meaning.
Release quality notes: Version 1.1 flags 40 marks with no time or frequency span (18 in test_soundscape and 22 in train_xenocanto); exclude them when training or checking event boundaries. It also documents one short audio file whose four labeled events extend past the file's end.
Rights: The record declares CC BY 4.0 for soundscape audio and all metadata/annotations. The Xeno-Canto subset keeps per-record terms; some license fields are empty. No files were acquired.

Reading location: Zenodo record 21821133, v1.1, published 2026-08-06; DOI 10.5281/zenodo.21821133; schema.json preview and Changes in v1.1. Prior version: Zenodo record 20762099.

What this does — and doesn’t — establish

The full collection is 49.7 GB. train_soundscape uses weak point-centered labels with a fixed +/-2.5 second window; many rows have no vocalization type. Its annotation CSV is 8,559,233 bytes, above the collection tool's current 5 MiB per-file cap, so its anonymized project-group semantics and row distribution remain unknown. The smaller 2,401,049-byte test_soundscape.csv contains public test labels and was not inspected; it is not a concealed locked test, and a separate untouched evaluation would be needed. The data are mostly from Germany with a smaller Greek subset and are not a parrot-specific context corpus. The record gives CC BY 4.0 for soundscape audio and all dataset metadata/annotations, while train_xenocanto retains per-record Xeno-Canto licenses; the schema says some of those license fields are empty. The schema also identifies 40 zero-area boxes that must be excluded from localization tasks and one truncated Xeno-Canto audio file whose annotations extend beyond its duration. No project/site leakage, source overlap, annotation reliability or source-media privacy audit has been performed.

What we can build on

Use a large bird collection to test whether software can find and organize labeled sounds. Keep these acoustic categories separate from independently observed behavior, a receiver's response and any claim of meaning. The best species for a later communication-focused study may still be a companion parrot, but CEB does not answer that question.

A practical starting point

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

Available starting point: Zenodo CEB v1.1 record 21821133 and its schema.json preview. No audio, annotation CSV, code or weights were acquired.

Reuse limits: The record declares CC BY 4.0 for soundscape audio, metadata and annotations. train_xenocanto keeps its per-record source license, and the schema notes some missing license values. Select only rows with verified rights; the record's top-level license does not clear every audio file for reuse.

Our proposed experiment: No media or annotation rows were downloaded or analyzed and no model was trained. The training CSV exceeds the current collection cap; test labels remain uninspected and are public rather than a concealed holdout. If a detector trial is justified, first review a bounded intake plan, then inspect group structure, exclude zero-area boxes, verify source overlap and reserve a separate untouched evaluation. Detection performance would not establish behavior, intent or meaning.

The next reading task

CEB remains a separate detection-engineering candidate, not the first communication-adjacent ML question. If the training-table group audit is still worth pursuing, first prepare a bounded, rights- and privacy-reviewed intake plan that fits an explicitly reviewed integrity limit; do not inspect the public test-label CSV as a locked evaluation. Then verify project-group meaning, source overlap and duplicates, exclude degenerate boxes, and obtain qualified bioacoustic review before proposing any model run. A separate untouched evaluation would be required for benchmark claims.

Go to the sources

Central European Birds (CEB) acoustic dataset, Zenodo v1.1 ↗
Official Zenodo record and data dictionary for strongly and weakly labeled bird soundscapes and Xeno-Canto focal recordings. Its API lists train_soundscape.csv at 8,559,233 bytes and test_soundscape.csv at 2,401,049 bytes; neither table's rows were inspected.
GeoNames country reference point for Germany ↗
The official GeoNames country feature, used only as a rounded country-scale atlas reference for the CEB soundscape recording coverage.
GeoNames country reference point for Greece ↗
The official GeoNames country feature, used only as a rounded country-scale atlas reference for the CEB soundscape recording coverage.

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.