OpenWhistle releases 180,000 dolphin whistles for self-supervised model training
Researchers introduce OpenWhistle, described as the largest public dataset of dolphin vocalizations, with a processing pipeline and benchmarks for whistle detection and classification.
In plain words
A team has published a large collection of dolphin whistle recordings so that machine learning models can be trained to analyze them. Self-supervised models learn patterns from raw audio without needing labels for every example. The dataset comes with expert-checked labels for a subset and standard tests to compare models fairly.
What was announced
The corpus contains roughly 180,000 whistles totaling 114 hours, recorded over five years from a stable pod of five bottlenose dolphins (Tursiops truncatus) in a semi-natural environment. A curated subset of 8,354 whistles carries expert annotations, and the release includes reproducible evaluation protocols for whistle-type detection and classification plus the full pipeline for detection, segmentation, and categorization. The authors position the resource against existing bioacoustic datasets they describe as mostly bird-focused and shallow per species, with dolphin data small, fragmented, and largely closed.
To show the dataset's usefulness, the team pretrained a Wav2Vec2.0 model adapted to dolphin acoustics on the corpus. They report that it learns effective representations and outperforms general-purpose bioacoustic models such as AVES and BioLingual on both benchmark tasks, while noting meaningful headroom remains for future work. According to the authors, this is the first open dolphin whistle dataset tailored for training self-supervised models.
What it does not show
- All claims come from the paper's own abstract; the benchmark results and the 'largest public dataset' status are author reports, not independently confirmed.
- The recordings cover only five individuals in one semi-natural pod over five years, so generalization to other populations or wild settings is not established by this evidence.
- The supplied material is an abstract excerpt only; it does not cover the full paper's methods or results.
For security and IT teams
- Scope
- The release covers a single pod of five bottlenose dolphins in a semi-natural environment, not wild or multi-population data.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.