Great tit (Parus major) / Single-population wild song dataset

The Wytham great-tit dataset maps song structure, not meaning

Peer-reviewed dataset paper; author dataset linked at OSF 2024 · Note reviewed 2026-10-25

The Wytham Great Tit Song Dataset is an unusually dense wild-population resource: its authors report 109,963 annotated songs and 1,161,033 acoustic notes from more than 400 individual birds, collected in Wytham Woods, Oxfordshire, from 2020 to 2023. The complete current OSF API root lists 23 files totaling 12,184,678,510 bytes: 14 data files (12,184,652,639 bytes) and nine documentation files (25,871 bytes). The four song-archive parts total 11,380,411,540 bytes and ten metadata/map files total 804,241,099 bytes. The OSF node declares CC BY 4.0 and reports last modification on 2023-07-10; no explicit dataset version label was present in the inspected node response. The 2025 paper already trained a ViT metric-learning model on this corpus, generated 384-dimensional embeddings for all 109,963 songs, and used them to analyze repertoire similarity and demographic drivers of song culture. It then used automated clustering plus manual song categorization; the reported 0.87 validation accuracy belongs to a separate ResNet50 validation procedure. The dataset supports valuable song-structure and population-methods work, but it has no independently observed interaction context, receiver response or label for what a song means.

What we reviewed · Primary 2024 dataset paper and 2025 full article/methods reviewed; author documentation, OSF API manifest and documented coordinate fields reviewed. On 2026-10-25, one 6,426-byte documentation file was fetched only to verify its current OSF SHA-256 and MD5; its text was not reviewed or redistributed. No audio archive acquisition or model reproduction.

Reviewed the 2024 publisher article and author dataset documentation for collection design, labels and package format; Oxford Research Archive rights record; setup-code license and pinned revision; and the full 2025 Current Biology article including STAR Methods on feature extraction, similarity, clustering and validation. The 2025 methods describe a ViT metric-learning model trained on song-type labels from this corpus, 384-dimensional embeddings calculated for all 109,963 songs, and downstream repertoire-similarity/cultural analyses; the article reports 0.87 validation accuracy for a separate ResNet50 procedure on a random subset. The reviewed method text does not establish an individual-held-out test for that validation. The complete 2026-10-24 OSF root traversal lists 23 current file entries totaling 12,184,678,510 bytes: 14 data files and nine documentation files. An earlier same-day 14-entry traversal had listed only the data/ folder and omitted the separate docs/ folder; it is superseded. All 23 normalized paths and byte sizes match the 2026-10-23 snapshot. The saved 2026-10-23 SHA-256 for docs/metadata/main.md is malformed at 63 characters. On 2026-10-25, the 6,426-byte file was retrieved solely for integrity verification; its SHA-256 and MD5 match the current OSF API. Its content was not read or redistributed. The OSF node license relationship declares CC BY 4.0. The author nestbox documentation specifies British National Grid X/Y coordinates and says these data are also included in main.csv. No coordinate values, audio/annotation or analysis-code payloads were accessed; the one documentation file was not read or redistributed; no model was run and no independent specialist reviewed this Talk2Nature note.

No independent specialist review has been received.

Study at a glance

Reported corpus: 109,963 songs and 1,161,033 annotated acoustic notes from more than 400 birds.
Study scope: One wild great-tit population at Wytham Woods, U.K.; data collected in 2020–2023.
Shared package: The complete OSF API root lists 23 files: four song archive parts totaling 11.38 GB, ten metadata/map files totaling 804 MB and nine documentation files totaling 26 KB. The approximately 9 TB original continuous soundscape archive is not included in the normal share. No payload was acquired.
Prior analysis: A 2025 paper trained a ViT metric-learning model on this corpus and used embeddings from all 109,963 songs for repertoire-similarity and cultural analyses. Do not treat those embeddings as an independent held-out baseline.
Useful questions: The corpus can support methods work on note boundaries and song sequences, but new identity/similarity/cultural claims must be distinguished from the published 2025 analysis.
Not labeled here: Independently observed interaction context, receiver response, comprehension or communicative meaning.
Separate rights: The OSF node declares CC BY 4.0 for the dataset record; the setup-code repository separately declares MIT. Nested provenance, 2025 analysis files and model weights were not reviewed.
Wildlife-location fields: Author documentation specifies British National Grid coordinates in nestboxes.csv and says they also appear in main.csv. No values or data files were inspected; keep public geography coarse and do not expose precise coordinates.

Reading location: Merino Recalde et al. (2024), Animal Behaviour 211:111–122, DOI 10.1016/j.anbehav.2024.02.008; OSF DOI 10.17605/OSF.IO/N8AC9; author documentation; setup code commit c96dc2d7ec0cc8bf6220f4aaf5dd6d9b941e5ef7; Merino Recalde et al. (2025), Current Biology, DOI 10.1016/j.cub.2025.02.016, STAR Methods sections on repertoire similarity, automated clustering and manual categorization.

What this does — and doesn’t — establish

This is one wild population and dawn-singing research system, so its labels do not establish transfer to other sites, species, household phones or general communication. Dataset documentation says bird identity associated with a breeding attempt can be uncertain or a proxy when the bird was not caught. The ordinary shared package is not the approximately 9 TB of original continuous soundscapes used in preparation. The authors’ nestbox metadata documentation specifies location coordinates in British National Grid and says the coordinates are also included in main.csv; the OSF manifest separately lists a Wytham perimeter shapefile. We inspected these documentation descriptions only, not coordinate values or file contents. Treat those fields as potentially sensitive wildlife-location information: keep public geography coarse and exclude precise coordinates from any future derivatives unless an appropriate wildlife-location review establishes a need and safe handling. The complete 2026-10-24 API tree has 23 entries; the earlier same-day 14-entry result was a partial traversal of data/ and omitted docs/. All 23 normalized paths and byte sizes match the historical snapshot. The 2026-10-23 local manifest stored a malformed 63-character SHA-256 for docs/metadata/main.md. On October 25, the current 6,426-byte file matched the OSF API SHA-256 and MD5; the cause of the malformed historical string is unknown. The file content was not reviewed, and no other provider hash was recomputed. Annotation values, grouping keys and dataset-level splits remain unaudited. The prior ViT embeddings were calculated across all songs from the same corpus, so treating them as a held-out-individual baseline for a new Talk2Nature benchmark would be an invalid shortcut; this is an evaluation-design inference from the published training scope. The 2025 paper already covers the obvious deep-embedding, individual-similarity and demographic song-culture tasks. No independent specialist assessed this note.

What we can build on

A large corpus can teach us to measure patterns in animal sound before we try to interpret them. This dataset is a candidate for note, sequence and repertoire questions, not a conversation model. Keep it separate from the Egyptian-fruit-bat context-prediction candidate and provisional household-parrot app cohort. Any future model result should use whole-individual and time-aware grouping, compare simple acoustic baselines, and state that it predicts acoustic structure rather than meaning.

A practical starting point

Metadata-only song-structure candidate; not admitted to Talk2Nature training or the first communication-adjacent benchmark.

Available starting point: Peer-reviewed 2024 dataset paper, full 2025 article/methods review, author documentation including the nestbox coordinate field descriptions, a dated OSF API manifest snapshot and pinned setup-code repository. No audio, data table, coordinate values, code, model or weights were acquired.

Reuse limits: The OSF node declares CC BY 4.0, consistent with the 2024 paper's statement for dataset input/output files; the separate setup repository declares MIT. Nested provenance and the 2025 analysis code/model rights were not reviewed. Article text, dataset, analysis code and model weights have separate rights.

Our proposed experiment: No files were analyzed and no model was run. The 2025 paper directly covers deep song embeddings, repertoire similarity and demographic culture; do not propose those as novel Talk2Nature results. Revisit only for a distinct, specialist-approved acoustic task with inspected labels, a leakage-safe individual/time split, a simple baseline and locked evaluation. Omit precise coordinates from derivatives unless an appropriate wildlife-location review establishes a need and safe handling. Acoustic structure or identity prediction would not establish receiver understanding or meaning.

The next reading task

Do not prioritize Wytham as a new Talk2Nature ML benchmark unless an independent avian-bioacoustics and evaluation reviewer identifies a useful task beyond the published embeddings, re-identification and cultural analyses. If a distinct acoustic-structure task is proposed, inspect the 2025 analysis split/code, define an individual/time-aware holdout and baseline, and decide whether any location fields are necessary. Keep precise coordinates out of maps and public derivatives unless an appropriate wildlife-location review establishes a need and safe handling. Do not present the corpus as a communication benchmark or run AutoResearch before shared readiness gates pass.

Go to the sources

A densely sampled and richly annotated acoustic data set from a wild bird population ↗
Merino Recalde et al. Animal Behaviour 211 (2024), 111–122. Peer-reviewed dataset paper describing more than 109,000 wild great-tit songs and over 1.1 million acoustic units from one U.K. population. Publisher abstract/data availability reviewed; complete methods not audited.
Wytham Great Tit Song Dataset and documentation ↗
Authors’ dataset website and documentation for song files, note annotations, bird/breeding-attempt IDs, metadata, package size, processing and nestbox-location fields. Supports a song-structure description, not a behavior or receiver-response label claim.
Author documentation for nestbox location fields ↗
Wytham Great Tit Song Dataset, OSF record ↗
The OSF node declares CC BY 4.0. A complete 2026-10-24 recursive root inventory lists 23 entries (14 under data/ and nine under docs/). The earlier same-day 14-entry result omitted docs/ because it traversed only data/. The 23 normalized paths and byte sizes match the October 23 snapshot; one docs/metadata/main.md provider digest differs at the same size. See the corrected manifest.
Wytham Great Tit dataset setup code, pinned main revision ↗
Author-maintained scripts for preparing the shared Wytham song dataset. GitHub declares MIT for this code repository; that license does not automatically apply to data or model weights.
The demographic drivers of cultural evolution in bird song: a multilevel study ↗
Merino Recalde et al., Current Biology (2025). The full article reports a ViT metric-learning analysis of all 109,963 Wytham songs for repertoire similarity and cultural-demographic questions. This is material prior art for any new embedding, re-identification or song-culture task.
Wytham Great Tit dataset public OSF JSON:API manifest ↗
Read-only JSON:API endpoint for the public OSF node and its osfstorage file tree. The complete 2026-10-24 root lists 23 files totaling 12,184,678,510 bytes (14 data files and nine documentation files) and the node-declared CC BY 4.0 license. One documentation-file provider digest differs from the prior snapshot at the same byte size. No payloads or location values were retrieved.
Wytham metadata-file integrity verification ↗
Integrity-only verification of one 6,426-byte dataset documentation file; no animal recording or annotation payload.

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.