Bottlenose dolphins / Longitudinal bioacoustic benchmark

OpenWhistle: a strong dolphin sound benchmark, not a translator

arXiv v1, September 2026; the authors’ pinned README lists it as a NeurIPS 2026 Datasets & Evaluations Spotlight. 2026 · Note reviewed 2026-10-24

OpenWhistle is a timely single-species research candidate: its authors report about 114 hours of 96 kHz recordings collected over five years from five known dolphins at Dolphin Reef in Eilat, plus 8,354 expert-annotated whistles in 10 categories. The authors describe the site as a semi-natural marine area open to the Gulf of Aqaba and say human–dolphin interactions are dolphin-initiated and voluntary. The official Hugging Face cards list 31,780 unlabeled audio segments and a separate labeled set split into 5,848 training, 1,253 validation and 1,253 test examples. Published tasks detect whistle presence or classify whistle type. These are upstream descriptions, not a Talk2Nature reproduction.

What we reviewed · Targeted paper, pinned dataset cards, official dataset license, and separate code/model rights review; metadata only

Reviewed the arXiv v1 paper data-collection, annotation, benchmark, evaluation, limitations and ethics sections; revision-pinned Hugging Face pretraining and classification dataset cards and file inventories; the separate revision-pinned model card; and the pinned official OpenWhistle README (source 411) and LICENSE-DATA (source 412) at commit ad715bb3cf24c36f85b3c3c8159077548134a39e. The repository README applies CC BY 4.0 to its four named dataset releases and LICENSE-DATA grants those terms; the dataset-card API license fields remain null. No audio, labels, code or weights were acquired or run. No independent specialist reviewed this note.

Not yet independently reviewed by a cetacean-bioacoustics or machine-learning specialist.

Study at a glance

Reported corpus: About 114 hours, five years and five known dolphins at one site; the labeled subset contains 8,354 whistles across 10 categories.
Published tasks: Whistle-type detection and classification, not receiver response, communicative function or meaning.
Evaluation boundary: The paper reports a 70/15/15 session-level split. It does not report a whole-individual held-out test.
Rights boundary: Official repository README and LICENSE-DATA at a pinned commit declare CC BY 4.0 for the datasets; Hugging Face card metadata is null. The separate model card says its license is unspecified, and no separate code license was visible. No files acquired.

Reading location: arXiv:2609.34839v1, Sections 3–5 and 9–10; OpenWhistle-Pretraining revision eea37423c7841b0f99a7c0380f945999c0a2c489; OpenWhistle-Classification-Finetuning revision 2b7806e3c5913c459c40b824ab7f944d494b3993; OpenWhistle-Wav2Vec2.0 revision 985ac11bc6939af5e41383143aad5a0ae6993f68; pinned official OpenWhistle README (source 411) and LICENSE-DATA (source 412) at commit ad715bb3cf24c36f85b3c3c8159077548134a39e.

What this does — and doesn’t — establish

All recordings come from one site and a pod of five; the paper says results need validation on independent groups. Its labels cover whistle categories, including individual signature-whistle types, not experimentally established message meanings or receiver understanding. The main evaluation holds out recording sessions, not whole animals, and the expert labels cover only a short part of the five-year span; the labels are highly imbalanced. The official repository README and LICENSE-DATA declare CC BY 4.0 for the dataset releases, while the pinned Hugging Face dataset-card fields are null; code terms were not visible in the repository root and the separate model card says its license is unspecified. No files are admitted or acquired.

What we can build on

This is a strong newly reviewed large-data candidate for a bounded communication-adjacent benchmark, not evidence of dolphin translation. A useful contribution would need to go beyond repeating the published whistle classifier—for example, a carefully reviewed test of temporal and individual generalization if the available metadata and rights support it. Keep this research track distinct from a companion-parrot app: app participation and model training have different evidence, consent and welfare requirements.

A practical starting point

Metadata-only candidate; not admitted to Talk2Nature training or public app inference.

Available starting point: Pinned arXiv paper, Hugging Face metadata/cards and official repository README/LICENSE-DATA only. The dataset release is declared CC BY 4.0 upstream; no audio, labels, code, checkpoint or video was acquired.

Reuse limits: The official repository commit ad715bb3cf24c36f85b3c3c8159077548134a39e explicitly declares CC BY 4.0 for the four listed OpenWhistle dataset releases; the Hugging Face card license fields are null. No separate code license was visible at the repository root, and the model card says the model license has not been specified. The dataset declaration does not cover code or weights; use attribution and verify exact scope before any acquisition or commercial use.

Our proposed experiment: No files were acquired or analyzed, no model was trained and no result was reproduced. Any future study needs independent review, a bounded approved run, a frozen session split, and a clear report of within-pod limits. Whistle labels alone cannot establish semantics or translation.

The next reading task

The dataset-level license declaration is now documented. Before any data acquisition or training, ask independent cetacean-bioacoustics and ML reviewers whether a within-pod whistle-structure reproduction would be useful and whether a temporal or individual-generalization extension is defensible. If they approve, freeze a bounded dataset-only run card and consider a project-authored simple acoustic baseline; do not reuse the unlicensed/unclear source code or model weights unless their terms are clarified. Do not infer meaning from detection or classification, and do not create playback features from this dataset.

Go to the sources

OpenWhistle dolphin dataset and benchmark paper (arXiv v1, 2026) ↗
Mustun et al., OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations, arXiv:2609.34839v1. The preprint reports acceptance as a NeurIPS 2026 Datasets & Evaluations Spotlight.
OpenWhistle pretraining audio dataset card and file inventory (revision eea3742) ↗
Official Hugging Face snapshot eea37423c7841b0f99a7c0380f945999c0a2c489 for the unlabeled 96 kHz dolphin-audio corpus. The pinned card lists 31,780 segments totaling 114.284 hours and a repository size of about 79.8 GB.
OpenWhistle expert-labeled classification dataset card and file inventory (revision 2b7806e) ↗
Official Hugging Face snapshot 2b7806e3c5913c459c40b824ab7f944d494b3993 for the 8,354-example, 10-category expert-annotated whistle set; it lists train, validation and test splits and a repository size of about 6.65 GB.
OpenWhistle Wav2Vec2 model card and license statement (revision 985ac11) ↗
Official Hugging Face snapshot 985ac11bc6939af5e41383143aad5a0ae6993f68 for the released dolphin-adapted Wav2Vec2 checkpoint. The pinned model card states that the model license has not been specified.
OpenWhistle author repository README at pinned commit ↗
The authors’ pinned README lists OpenWhistle as a NeurIPS 2026 Datasets & Evaluations Spotlight, identifies four dataset releases, and distinguishes dataset licensing from the repository code.
OpenWhistle dataset license at pinned repository commit (LICENSE-DATA) ↗
The authors’ LICENSE-DATA file grants CC BY 4.0 for the four dataset releases named by the pinned README. It does not grant a code or model license.
GeoNames populated-place reference for Eilat, Israel ↗
Official GeoNames city point used only as a rounded one-decimal atlas marker for the OpenWhistle study region; it is not a recording or animal coordinate.

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.