Talk2NatureCross-species bioacoustics / Annotation software
Whombat: learn from an existing bioacoustic annotation workbench
Methods in Ecology and Evolution 16:19–28; version of record online 2024 2025 · Note reviewed 2026-10-13
Whombat is a browser-based workbench for managing bioacoustic projects, reviewing recordings and human annotations, and visualizing or exchanging machine-learning predictions. The paper describes bat and avian applications. Its guide imports an AOEF dataset with an audio directory first, then its annotation project; evaluation sets and model runs are separate, and predictions are distinct from human annotations. Talk2Nature now offers an optional experimental one-way AOEF 1.1 package. The exact export imported into Whombat 0.9.0 with one recording, intervals [1, 2] and [4, 5] seconds, four visible namespaced project tags, and two notes. A missing AOEF project_tags list and a macOS canonical-path issue were identified and fixed/documented during this synthetic compatibility test.
What we reviewed · Article and guides reviewed; synthetic AOEF 1.1 export loaded with soundevent 2.10.0 and imported through Whombat 0.9.0 desktop UI; specialist review pending
Reviewed the complete primary article through the publisher/UCL open-access record and the pinned Whombat README/license and PyPI release metadata. Read the official import, dataset, annotation-project, evaluation and installation guides. Verified the temporary Whombat 0.9.0 wheel SHA-256 against PyPI and inspected the separate Whombat GPL v3 and soundevent MIT licenses. In an isolated Python 3.12.5 environment on macOS 26.4, the exact current Listen export passed Whombat's dataset-then-project UI import. The app created one recording, two TimeInterval annotations, two notes, one task and four project tags; checksum, sample rate and duration matched, the tag table showed both sound events, and both imported creator IDs remained null. Core resolved versions were FastAPI 0.112.2, fastapi-users 12.1.3, Starlette 0.38.6 and SQLAlchemy 2.0.54. The test used only an invented WAV and a temporary SQLite database; the wheel, environment and test outputs were removed. No animal audio, participant record or model weight was tested.
Not yet independently reviewed by a bioacoustics or research-software specialist.
What this does — and doesn’t — establish
This is annotation and model-assisted review infrastructure, not a decoder of animal intent or a general animal-language translator. The app-level check establishes format interoperability only for Whombat 0.9.0 and the recorded test stack; other dependency versions, mobile behavior, real recordings, label quality and research suitability remain unverified. The test used invented audio only. Full Talk2Nature IDs, declared animal context and any Station journal stay in an unencrypted provenance sidecar; inspect it before sharing. Whombat's path check requires one canonical absolute audio-root path; on macOS /tmp may resolve to /private/tmp. No reverse import exists, and the sidecar is still the source of truth. The Whombat v0.9.0 code is GPL v3; the soundevent library is MIT; the article is CC BY 4.0. These licenses cover separate assets.
What we can build on
Do not build a second full research annotation editor before proving a distinct user need. Keep the novice field companion focused on bounded observation and local review; investigate an optional research-workbench exchange path, with model suggestions preserved separately from accepted human labels.
A practical starting point
Evaluate as a separate interoperability candidate
Available starting point: Whombat v0.9.0 annotation and model-prediction exchange workflow
Reuse limits: Article: CC BY 4.0. Pinned source code: GNU GPL v3. Sample audio, dependencies, and model artifacts have separate terms. No code or sample assets copied.
Our proposed experiment: The synthetic export passes soundevent 2.10.0 save/load and Whombat 0.9.0 desktop UI import with one recording, two intervals, four project tags and two notes. Human labels remain separate from model predictions and the original record remains in a sidecar. Next obtain specialist review; do not infer biological meaning from an exchange test.
See the related tools and reading paths →
The next reading task
Have an independent bioacoustics or research-software specialist review the AOEF field mapping before recommending it for a study. Consider other Whombat/dependency versions only in a fresh synthetic test. Keep the feature experimental; real recordings, participant intake and model-training admission require separate rights, consent, privacy and scientific-protocol gates.
Go to the sources
- Whombat: peer-reviewed audio annotation workflow ↗
Martínez Balvanera et al. (2025), Methods in Ecology and Evolution 16:19–28. The article describes audio project management, human annotation/review, visualization, and import/export of model predictions. It is a tool-method reference, not evidence of animal communication or a vetted Talk2Nature integration.
- Whombat v0.9.0 source and release provenance ↗
Pinned source commit 6cc44a9ad788087ca5bfdd6cf6f22f464b94748e, published as v0.9.0 on PyPI on 13 May 2026. PyPI reports Python >=3.12, a source-distribution SHA-256 of dbb653fb94a7221134699dcb2a1ce5e28f63c1f86ae9ae7e34bd5adb214d7c79, and a verified GitHub Actions publishing attestation. The checksum is provider-supplied and was not locally verified because no package was downloaded. The pinned README and GNU GPL version 3 license were read.
- Whombat official import, annotation, evaluation and privacy guides ↗
The guide requires importing an AOEF dataset with an audio directory before importing the annotation project. Evaluation sets and model runs are separate, and predictions are distinct from human annotations. The installation guide notes exported files may include user identity.
- soundevent 2.10.0 and Audio Object Exchange Format ↗
The official MIT-licensed library supports AOEF v1.1.0 dataset and annotation-project serialization. A version-pinned wheel's SHA-256 matched PyPI metadata; save/load was tested on an invented WAV and annotations only.
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.