Best large-data path for learning song-structure methods. The labels describe songs and notes; they do not explain what a call means.
Research details
The shared, prepared dataset describes 109,963 songs, more than 1.16 million annotated acoustic units and over 400 birds in one U.K. population. The authors say the original roughly 9 TB of continuous soundscapes are not included. A 2025 study already used deep metric learning across the corpus to analyze repertoire similarity and song culture. Any new work must be a scoped reproduction or a clearly distinct question.
The authors' metadata guide documents precise British National Grid nestbox coordinates in two tables, plus a study-perimeter file. We have not inspected coordinate values or acquired data files. Treat these fields as potentially sensitive and keep public maps at coarse study-region scale.
A strong single-species resource for testing whistle detection and expert-labeled whistle categories. Those tasks do not decode messages.
Research details
The preprint reports about 114 hours of audio across five years from five known dolphins, with 8,354 expert-labeled whistles. Its published evaluation separates recording sessions but keeps the same pod across partitions, so it does not establish generalization to new dolphins.
The official repository declares CC BY 4.0 for its datasets, though Hugging Face cards omit a license field. The code has no separate visible license, and the Wav2Vec model license is unspecified. No media has been acquired; independent cetacean and ML review is still required.
A rare dataset connects a well-defined movement signal with follower behavior and feeder outcomes. The core dance–food relationship is established prior work, not an undiscovered animal language.
Research details
A 2023 study combines waggle-phase detection, individual tracking, follower behavior and feeder observations. The Zenodo record lists more than 46 million 2019 track detections in an approximately 4.9 GB release; the paper also reports more than 100,000 waggle phases and an annotated subset spanning 4,389 bee IDs. Dataset metadata declares CC BY 4.0, while code and model terms are separate.
The recordings come from one Berlin observation-hive site. The published classifier groups sample windows by hour rather than holding out bees or colonies, so it does not establish generalization to new bees or hives. The source includes fine-scale feeder coordinates; this project keeps them out of public summaries. No dataset files were acquired or run.
A promising bird-specific reading path connects machine call-label predictions with a controlled test of how finches distinguish labeled calls. The underlying sound-to-trial relationship still needs verification.
Research details
A 2025 study tested 12 adult finches on 11 researcher-defined call types. A separate 2026 paper reports an 11-type classifier on an archive described as 3,433 recordings. The classifier predicts supplied labels; it does not translate calls. The papers and dataset descriptions differ on bird counts and whether files contain isolated calls or sequences.
Before a distinct follow-up, reconcile file versions, bird identities and unitization, resolve component rights, and determine whether the classifier recordings can be joined to the behavioral trials. Talk2Nature has not acquired the archive or reproduced either result.
The closest reviewed lead for studying turn-taking and measurable responses in controlled vocal exchanges—not a translated conversation.
Research details
A 2026 preprint reports extensive vocal-interaction experiments and changes in birds’ calling during interactive versus non-interactive playback. The separate data deposit has conflicting CC BY-SA and CC BY-NC declarations. Four raw example tracks do not establish that the complete analyzed corpus is available, and bird-level independence across splits remains unverified.
A published context-prediction question, but its current connected groups do not support a balanced held-out evaluation.
Research details
A 2024 paper reports predicting video-coded interaction context from 35,074 vocalizations across 11 classes. These are published results, not a Talk2Nature reproduction, and do not show that bats understood a message.
The recordings used a specialized 250 kHz laboratory system, not ordinary phones. A metadata-only reconstruction matches the reported filtered count, not the exact rows or folds. Keeping each file and every positive known participant together leaves one group with 95.6% of examples. Specialist review and additional session/context metadata are needed; we have not trained a model.
A practical candidate for testing the phone workflow with existing carers; the app records observations and does not identify the animal.
Research details
Budgerigars have published research on learned contact calls and social interaction. Available public data are not a ready-made, event-aligned phone-audio-to-behavior training set.
A household study needs multiple independent homes, an avian-reviewed behavior codebook, consent and privacy safeguards, and a retention and withdrawal plan. No cohort has been recruited.
V33DA is a narrow caller-attribution task, not a communication test.
Research details
V33DA reports 33,625 physically verified call events from 10 zebra finches. Its labels support caller attribution among visible birds, not behavior, receiver response or call meaning. A read-only inventory of its pinned 100-event code sample lists 334 files totaling 62,503,976 bytes; the sample-specific split and several metadata joins remain unverified.
This is a candidate for a small methods exercise only if independent avian and ML reviewers approve the question and grouping. No sample data or model has been acquired or run.
Large bird and multi-species archives can help test whether software detects sound events or recognizes annotated species and call types. Those labels do not tell us what an animal meant. The new Northeastern forest-bird release is a promising field reference for detection, but its paper abstract and latest data version disagree on the number of species. Central European Birds is another detection candidate with a different label and rights profile. Neither is a behavior or receiver-response dataset. Dogs, cats, frogs and other animals remain welcome in the observation tools, without a claim that the app identifies every species.
For a large-data methods exercise: can a model reproduce expert-marked song-note boundaries on Great Tit recordings from birds it has not seen? That tests an observable acoustic task, not meaning, and would need to be checked against the published methods before it is treated as new research. A separate future question is whether calls predict independently video-coded social context; the current bat candidate does not yet support a balanced grouped holdout.
For either task, freeze the labels and test set, keep related animals and sessions together, compare simple baselines, and have a qualified reviewer assess the design before training.
A prediction is a candidate association. A separate, ethically reviewed study measuring a receiver’s response would be needed to test a bounded communication claim. The Field Companion uses passive, user-started observation; it does not play calls to animals.
The Field Companion welcomes observations of many animal groups. Species choices are entered by the observer; the app does not identify an animal. Try the invented, microphone-free rehearsal first, or read the research plan before contributing to any future study.
Updated October 25, 2026. Evidence-led candidate choices only: no Talk2Nature animal cohort, independent specialist review, model training or translation result is claimed. Real-data intake to the private research database remains closed.
Talk2Nature