Talk2NatureSpotted hyenas / Preprint and sound-event dataset
HyenaSET: a strong sound-event benchmark, not a communication test
bioRxiv v2 preprint; Edmond dataset v1 2026 · Note reviewed 2026-10-13
HyenaSET is a useful recent sound-event methods comparator: its Edmond v1 metadata describes 590,480 ten-second clips, about 1,640 hours from 19 spotted hyenas, with 87,394 expert-reviewed clips (about 243 hours). The annotations mark vocalization intervals and types, and distinguish calls from the collar-wearing focal hyena from nearby non-focal callers. The preprint provides sample-stratified splits and a four-individual held-out split, then reports an animal2vec baseline. These targets concern event detection, call categories and source attribution; they are not independently observed receiver responses or evidence of message meaning.
What we reviewed · Targeted preprint methods and results; official dataset version, file and terms metadata
Read the bioRxiv v2 abstract and targeted data-collection, annotation, split, benchmark and data/code-availability sections. Examined official Edmond and DataCite metadata for version 1, release date, archive size, provider checksum and dataset terms. The paper is a preprint, not peer-reviewed. No audio, labels, code or model weights were downloaded, installed or run; no data rows, coordinates or individual timestamps were inspected. No independent review or reproduction occurred.
No independent bioacoustics or machine-learning specialist has reviewed this Talk2Nature assessment.
Study at a glance
- Recording inventory
- Edmond v1 metadata lists 590,480 ten-second clips (about 1,640 hours) from 19 spotted hyenas; the archive is about 107 GB.
- Reviewed labels
- 87,394 clips (about 243 hours) were expert-reviewed; 10,293 clips contain at least one marked focal or non-focal vocalization. Labels are sound events, not receiver responses.
- Evaluation splits
- The dataset description lists five sample-stratified splits and a four-individual held-out split. The preprint says the animal2vec baseline is pretrained on all available audio before fine-tuning on the first split.
- Rights and reported result
- Dataset: CC BY-NC-SA 4.0. Preprint: CC BY 4.0. Code and weights are separate. The preprint reports a micro-average AP of 0.86; Talk2Nature did not reproduce it.
Reading location: Woerner et al., bioRxiv v2: abstract and targeted sections on data collection, annotation, train/test splits, animal2vec pretraining, benchmark table and data/code availability. Edmond v1: official dataset and API metadata for release date, size, checksum and custom terms. Checked 2026-10-13; no data files acquired.
What this does — and doesn’t — establish
The bioRxiv manuscript is not peer-reviewed. Edmond's v1 archive is one 107,184,528,475-byte TAR.GZ with a provider MD5 that Talk2Nature has not locally verified. Its dataset-specific terms are CC BY-NC-SA 4.0, so the data are not admitted to the shared commercial-capable training path; the preprint, animal2vec code and model weights have separate rights. The record describes a CSV that maps randomized names to individual identities and recording times for four test individuals; no rows or precise locations were reviewed, so wildlife-location sensitivity remains unresolved. The preprint says animal2vec is pretrained on all available audio before fine-tuning on the first cross-validation split. That wording implies exposure to audio from held-out partitions; our methods interpretation is that the reported result is transductive unless the authors clarify otherwise, not a strict unseen-audio baseline. No independently observed behavior or receiver-response labels were established by this review.
What we can build on
A useful design reference can still be unusable as project training data. Preserve event boundaries, focal/non-focal attribution, quiet intervals, animal-held-out tests and a separate record of what data a representation model sees during pretraining. A held-out label split is not a clean generalization test if unlabeled test audio entered pretraining. Keep signal detection and source attribution separate from independently observed context, receiver response and meaning.
A practical starting point
Metadata-only methods comparator; not admitted to the shared commercial-capable data path
Available starting point: bioRxiv preprint (CC BY 4.0); Edmond dataset v1 (CC BY-NC-SA 4.0); animal2vec code and checkpoints are separate assets
Reuse limits: The dataset license is non-commercial and share-alike. Its public availability does not authorize commercial use in Talk2Nature. No data, code or weights were acquired.
Our proposed experiment: No Talk2Nature model run is authorized by this review. A future separately governed non-commercial reproduction would still need rights and privacy checks, reviewer approval, split-safe pretraining and exact baseline reproduction.
See the related tools and reading paths →
The next reading task
Keep the records metadata-only. Before considering a separate non-commercial reproduction, have a qualified bioacoustics and machine-learning reviewer inspect the exact label dictionary, group manifests, pretraining exposure and wildlife metadata; first establish a concrete question that this 107 GB collar-recorded dataset can answer. Do not acquire it or use its weights in the shared project path without a scoped rights, privacy and evaluation review.
Go to the sources
- animal2vec code repository and archived release ↗
Official repository at github.com/livingingroups/animal2vec; the 2026 article states MIT and cites archived release DOI 10.5281/zenodo.17640321. No commit-pinned code snapshot or runtime was inspected.
- HyenaSET: Hyena Sound Event Transcripts preprint v2 ↗
Woerner et al.'s bioRxiv preprint describes a large spotted-hyena sound-event dataset and an animal2vec benchmark.
- HyenaSET dataset v1, Edmond ↗
Official Edmond dataset DOI for HyenaSET version 1, published June 22, 2026.
- HyenaSET v1 terms and file metadata API ↗
Official Edmond API metadata declares the dataset-specific CC BY-NC-SA 4.0 terms and the v1 archive inventory.
- GeoNames country reference point for Kenya ↗
The official GeoNames country point for Kenya, used only as a rounded country-scale atlas reference for literature study regions.
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.