Research map / Label-efficient and robust learning / Self-supervised and contrastive
OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces OpenWhistle, a five-year, ~114-hour dolphin whistle corpus with an expert-annotated benchmark for whistle detection and type classification, and shows a Wav2Vec2.0 model pretrained in-domain beats general bioacoustic models under linear probing.
Problem
Existing bioacoustic corpora are bird-centric and shallow per species, and dolphin whistle datasets are small, fragmented, and largely closed, which limits self-supervised and longitudinal study of dolphin communication.
Approach
A large, open, longitudinal recording corpus from a stable known pod, paired with an expert-labeled whistle-type subset and two linear-probe benchmark tasks, plus an in-domain Wav2Vec2.0 pretrained from raw audio on that corpus.
Key results
- Wav2Vec2.0 pretrained on OpenWhistle reaches 81.1% accuracy on whistle-type classification and 75.6 mAP on detection under linear probing.
- Best off-the-shelf model AVES-bio reaches 75.1% classification accuracy and 65.0 mAP detection; BioLingual 71.3% and 66.5 mAP.
- Whistle presence CNN achieves 96.52% precision and 97.99% recall on a held-out test set.
Surprising findings
- BioLingual has higher detection mAP (66.5) than AVES-bio (65.0) despite lower classification accuracy, so rankings differ across tasks.
Limitations
- Single site and five-individual pod limits diversity and generalization.
- Uneven temporal coverage with a 2022 gap; expert subset spans only 5 months.
- Highly imbalanced expert labels with some classes under 100 examples; detector recall gaps may miss whistles.
Experimental setup
Datasets: OpenWhistle pretraining corpus OpenWhistle expert-annotated subset WMMSD DCLDE 2011
Benchmarks: OpenWhistle whistle-type classification OpenWhistle whistle-type detection
Metrics: accuracy mean average precision F1
Baselines: AVES-core AVES-bio BioLingual Mean spectrogram MFCCs Spectral features
Builds on: Wav2Vec2.0 VGG16 ARTwarp CREPE AVES
Closest papers by meaning
- NatureLM-audio: an Audio-Language Foundation Model for Bioacoustics (ICLR 2025 · similarity 0.89)
- Towards A Translative Model of Sperm Whale Vocalization (NeurIPS 2025 · similarity 0.87)
- Vocal Call Locator Benchmark (VCL) for localizing rodent vocalizations from multi-channel audio (NeurIPS 2024 · similarity 0.86)
- The iNaturalist Sounds Dataset (NeurIPS 2024 · similarity 0.85)
- BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics (ICLR 2025 · similarity 0.84)
- Hear You Say You: An Efficient Framework for Marine Mammal Sounds’ Classification (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.