Birds, mammals, amphibians and insects / Passive acoustic dataset and benchmark

El Silencio helps detect sound across taxa, not decode meaning

Dataset publication year 2025 · Note reviewed 2026-10-22

El Silencio is a useful multi-taxon sound-event detection candidate: the paper describes 770 one-minute soundscapes (12.8 hours) and about 15,000 event boxes across birds, mammals, amphibians and insects. The manual protocol labels acoustic group, call/sonotype and sound form; it groups same-species signals separated by no more than one second. These acoustic/taxonomic labels do not supply independently observed behavior, caller intent or receiver-response labels. The paper's BirdCLEF+ competition dataset is a separate sound-identification task, so its classification score is not a result on this El Silencio subset and does not measure communication. This can help test a multi-species detection/data workflow, but it is not a communication or meaning benchmark.

What we reviewed · Full overview paper, pinned annotation protocol and dataset/benchmark-fit review

Queried official Zenodo API metadata for labeled record 17064353 and unlabeled record 17122738, including publication dates, record-level licenses, file counts, byte totals and provider MD5 values. Read the dataset repository README and MIT LICENSE at commit 5e2f7826c908fec3018740068bbdeda6c6977f35; reviewed all extractable Spanish text on pages 2-18 of its pinned protocol and the full 11-page CEUR overview paper. Local review copies of only the paper and protocol are excluded from the public site. No dataset audio, label table, code or model weights were acquired or run; no baseline was reproduced.

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

Study at a glance

Labeled release: Zenodo record 17064353; published 2025-09-05; 9 provider-listed files totaling 3,546,609,904 bytes; record-level CC BY 4.0.
Collection: 770 one-minute AudioMoth soundscapes, totaling 12.8 hours, from seven sites in the Middle Magdalena region; the repository describes dawn and dusk recordings from 2023.
Labels: About 15,000 strong acoustic-event boxes across birds, mammals, amphibians and insects. UNKNOWN is not a stable species class.
Annotation method: The Spanish protocol describes manual RavenPro Group/Label/SongType annotations. Same-species signals separated by no more than one second are grouped. The protocol's visible group codes cover birds, anurans and insects; alignment to the four taxa in the paper needs expert review.
Separate unlabeled release: Zenodo record 17122738; published 2025-09-15; 61 ZIP files totaling 49,620,235,222 bytes; CC BY 4.0. The pinned README describes 10,980 soundscapes at 61 sites.
Count reconciliation: The overview paper reports 11,340 unlabeled soundscapes at 63 sites. The current repository and separate unlabeled Zenodo release report 10,980 at 61 sites; the reason for the difference is not documented in the sources reviewed.
Task fit: A multi-taxonomic event-detection and species/sonotype-identification candidate, not independently observed behavior, caller intent, receiver response or semantic translation. The BirdCLEF+ competition score belongs to a separate task and dataset.

Reading location: Zenodo records 17064353 and 17122738, API metadata queried 2026-10-22; pinned repository README, LICENSE and Protocol.pdf at commit 5e2f7826c908fec3018740068bbdeda6c6977f35; full CEUR-WS Vol. 4038 paper 232.

What this does — and doesn’t — establish

The paper's unlabeled snapshot is 11,340 recordings at 63 sites; the later pinned README and separate Zenodo deposit describe 10,980 recordings at 61 sites (360 fewer across two sites), with no explanation in reviewed sources. The protocol's explicit group codes are birds, anurans and insects, while the paper describes four taxa including mammals; the mapping needs specialist clarification. No inter-annotator reliability statistic was reported in the reviewed paper/protocol. The labeled Zenodo page displays 109.8 GB while its nine-file API inventory totals 3,546,609,904 bytes; unresolved. The unlabeled API inventory lists 61 ZIPs totaling 49,620,235,222 bytes. Zenodo declares CC BY 4.0 at record level; source-media rights, consent/privacy, human-audio handling and sensitive-location conditions remain unreviewed. UNKNOWN is not a stable class. No audio, annotations or split were inspected, and no baseline was run. Event detection, acoustic association and communication meaning remain different claims.

What we can build on

Use multi-taxonomic soundscapes to check whether the tool can locate and organize candidate events across animal groups. Keep a human-reviewed event label separate from behavior observed in context, model predictions and any claim about meaning. A pretrained encoder may organize acoustic patterns; grounded communication claims still require independent behavioral observations and receiver-side tests.

A practical starting point

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

Available starting point: Zenodo labeled record 17064353 and separate unlabeled record 17122738, plus the pinned dataset README/license, Spanish protocol and 2025 overview paper. Paper/protocol review copies remain local and are not public build assets. No recordings, labels, model weights or code were acquired.

Reuse limits: The two Zenodo API records declare CC BY 4.0; the paper is CC BY 4.0 and the pinned repository declares MIT. Each applies only to its stated asset. Verify source-media rights, consent, privacy, human-audio and sensitive-location conditions before dataset use; neither document license resolves those questions.

Our proposed experiment: No data were downloaded or inspected, no code was run and no model was trained. Any future detection baseline must use a documented, rights-cleared release and site-grouped evaluation. It cannot satisfy the independent behavior-label or receiver-response requirements for a communication claim.

The next reading task

Ask a qualified bioacoustics reviewer to resolve the protocol's taxonomic-label scope, annotation consistency and appropriate use. Seek provider clarification for the 11,340/63 versus 10,980/61 release change and the labeled record page/API size discrepancy; review source-media rights, privacy, human-audio and location safeguards before acquiring any sample. If a concrete detector question remains, pre-register a small rights-cleared evaluation with site-held-out splits and a simple baseline before running a model. This dataset does not satisfy the first context/meaning benchmark gates.

Go to the sources

El Silencio labeled multi-taxonomic dataset, Zenodo record ↗
Version 1 deposit of 770 one-minute multi-taxonomic soundscapes and strong acoustic-event labels from the Middle Magdalena region, Colombia.
El Silencio unlabeled soundscapes, Zenodo record ↗
Separate Zenodo deposit linked by the dataset repository for the broader, unlabeled soundscape collection.
El Silencio dataset repository at a pinned GitHub commit ↗
Dataset README and license pinned to commit 5e2f7826c908fec3018740068bbdeda6c6977f35.
BirdCLEF+ 2025 multi-taxonomic overview paper ↗
Conference overview of the BirdCLEF+ 2025 sound-identification challenge and El Silencio dataset.
El Silencio manual multi-taxon annotation protocol ↗
Spanish-language Protocol.pdf for manual RavenPro annotation, pinned to the dataset repository commit.
GeoNames country reference point for Colombia ↗
The official GeoNames country point for Colombia, used only as a rounded country-scale atlas reference for the El Silencio study region.

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.