voxfauna
FaunaCore v0.1 · pre-release · open specification
Xeno-canto holds more than a million recordings across roughly 12,900 species. iNaturalist has passed a million more. Nearly every one of them carries a single label: which species made this sound.
That trains a species classifier. It does not train meaning. Meaning is not in the sound — it is in the function of the sound in a context. So VoxFauna records the two things every existing corpus throws away: what happened immediately before, and what the animal did next.
the gap
106
Xeno-canto: 1,000,000+ recordings, ~12,900 species. iNaturalist: 1,000,000+ sound observations from 27,000+ recordists. Both are labelled at the level of species, not meaning. They are the raw material. They are not the answer.
11
The 2026 Coller-Dolittle Prize went to Dr Julie Elie for decoding 11 core zebra finch calls — distress, hunger, greeting — and showing the birds sort them by meaning, not by acoustics. It took over a decade of manual context annotation. On one species. That decade is the bottleneck.
14,500
Google DeepMind's Perch 2.0 was trained on 1.5M+ recordings across ~14,500 species. Earth Species Project's NatureLM-audio is open source and can already separate distress from affiliative calls. The models are here, and they are free. What they are starving for is labelled context.
“The main thing stopping us cracking the code of animal communication is a lack of data.”
— Jonathan Birch, LSE · Coller-Dolittle scientific panel
the schema
FaunaCore is an open specification for a single unit of animal communication data. It is deliberately boring, deliberately relational, and deliberately compatible with Darwin Core and Audiovisual Core, so that anything captured here can be published straight to GBIF and ingested by anyone.
A — antecedent
What triggered it. A van in the driveway. A hand near the feed bin. A hawk overhead. Coded, timestamped, and captured from the thirty seconds before the human ever reached for the phone.
B — signal
The vocalisation, and the body that produced it. Segment bounds, frequency bounds, posture, ears, tail. Multimodal, because most meaning is.
C — consequence
What the animal did next. Whether it repeated the signal. Whether anything resolved it. This is the field that turns a recording into evidence about function.
// faunacore-0.1 — abridged
{
"subject": { "taxon": "Canis familiaris", "known_individual": true },
"antecedent": {
"coded_terms": ["stimulus.vehicle_approach", "stimulus.stranger_present"],
"seconds_before": 4
},
"signal": {
"segment_ms": [29840, 31120],
"signal_type": "bark.alert",
"body": ["posture.tail_high_stiff", "posture.ears_forward"]
},
"consequence": {
"next_action": "action.approach_stimulus",
"resolved": true,
"seconds_to_resolution": 22
},
"interpretation": {
"meaning": "alert.notify_group",
"basis": "repeated_observation",
"contributor_certainty": "high"
},
"verification": { "status": "pending", "inter_rater_alpha": null },
"ethics": { "provocation_used": false }
}The specification is open, versioned, and unfinished. It is almost certainly wrong somewhere. That is the point of publishing it now rather than later.
what this is not
There is a wave of products promising to translate your pet, quoting accuracy figures with no published methodology and no independent validation. They will take a lot of money and then take a lot of credibility with them. This project is built to be the opposite of that.
No speech bubbles. No "your dog is saying…". We produce the data that would make decoding possible, and we let scientists do the decoding.
Perch, NatureLM-audio and BirdNET already exist and are open. Rebuilding them would be vanity. We are upstream of them.
Generating signals to influence animal behaviour is the highest-risk capability in this field. It is explicitly out of scope.
provocation_used: false is a required field, and anything else is rejected at ingest. A data platform must never create an incentive to distress an animal.
Every entry carries confidence, sample size, and inter-rater agreement. Where annotators disagree, you will see that they disagreed. Uncertainty is data, not embarrassment.
contact
If you work on this — in a lab, on a bioacoustics team, at a museum, in a shelter, in a training yard, or in a barn — I would rather hear your objection now than discover it after fifty thousand records have been collected in the wrong shape.
Best for detailed feedback, a collaboration proposal, or if you want to contribute data or review. Labs and teams can also register interest to participate →