OpenWhistle

A Large-Scale Longitudinal Dataset and Benchmark
of Bottlenose Dolphin Vocalizations

An open, longitudinal resource for learning the fine-grained structure of dolphin communication.

Whistle sequencesOpenWhistle representationsAcoustic structure

≈180KWhistles

114hAudio

5 yrsLongitudinal span

8,354Expert labels

Read paper View code Explore data NeurIPS 2026 Spotlight

01 / THE GAP

Bioacoustic datasets go broad.
OpenWhistle goes deep.

Most animal-audio corpora cover many species with only a shallow sample of each. That is useful for species recognition, but not enough to study one communication system over time.

OpenWhistle follows a stable, known pod in a semi-natural marine environment—preserving the scale, continuity and acoustic messiness needed for new questions.

02 / THE DATASET

One species.
Unprecedented depth.

180,000

whistles

The largest publicly available dolphin-vocalization corpus to date.

114

hours of audio

Raw acoustic material suited to large-scale self-supervised pretraining.

years

Longitudinal recordings with contiguous sequences and changing social context.

8,354

expert labels

Whistles labeled by experts within their original sequence context.

Aerial view of Dolphin Reef showing the fixed hydrophone positions — Recording site
Fixed hydrophones follow the pod continuously in its everyday environment at Dolphin Reef, Eilat.

Dolphin Reef lagoon with hydrophone positions and a dolphin visible underwater — Recording site
Fixed hydrophones follow the pod continuously in its everyday environment at Dolphin Reef, Eilat.

03 / THE REPERTOIRE

Every contour
has a character.

Seven individually distinctive signature-whistle types and three non-signature types form the expert-labeled repertoire. Their contours remain recognizable through the noise, overlap and variation preserved by OpenWhistle.

01–05 / 10

SW / 01LunaSignature whistle

SW / 02NeoSignature whistle

Spectrogram of Nana's rising signature whistle

SW / 03NanaSignature whistle

Spectrogram of Nikita's signature whistle

SW / 04NikitaSignature whistle

Spectrogram of Yosefa's signature whistle

SW / 05YosefaSignature whistle

SW / 06DanaSignature whistle

SW / 07ShySignature whistle

Spectrogram of non-signature whistle type 1

NSW / 01NSW 1Non-signature whistle

Spectrogram of non-signature whistle type 2

NSW / 02NSW 2Non-signature whistle

Spectrogram of non-signature whistle type 3

NSW / 03NSW 3Non-signature whistle

04 / THE PIPELINE

One pipeline.
Two datasets.

OpenWhistle dataset construction pipeline from raw audio through CNN detection and segmentation to the pretraining and expert-annotated datasets

05 / THE BENCHMARK

In-domain listening
makes the difference.

One dataset, four views of the results. Compare general audio models with OpenWhistle pretraining, data scale and end-to-end fine-tuning.

OpenWhistle performance summary

64.37%Full-set macro-F1

+11.68 ptsvs best general backbone

Backbone	Full set10 types · F1	Natural6 types · F1	Balanced6 types · F1	DetectionmAP
AVES-Coregeneral audio	51.57 ± 2.34	51.00 ± 2.68	67.39 ± 2.14	57.40 ± 2.10
BioLingualaudio–text	54.00 ± 2.77	52.93 ± 2.11	70.98 ± 2.05	66.50 ± 2.20
AVES-Bioanimal vocalizations	52.69 ± 2.31	51.24 ± 2.52	75.13 ± 2.03	65.00 ± 2.30
Wav2Vec2.0OpenWhistle · released checkpoint	64.37 ± 2.67	59.91 ± 2.57	81.67 ± 1.80	75.76 ± 2.02

Task

Linear-probe classification & detection

Classify expert-labeled whistle types with a frozen encoder, then detect whistle-bearing windows.

Classification: Macro-F1
Detection: mAP

Best continued-pretraining result

63.53%Full-set macro-F1

+10.84 ptsvs off-the-shelf AVES-Bio

Checkpoint	Full set10 types · F1	Natural6 types · F1	Balanced6 types · F1	DetectionmAP
AVES-Bio + OW10% of data	47.99 ± 2.32	50.03 ± 2.66	70.40 ± 2.14	65.43 ± 2.21
AVES-Bio + OW50% of data	50.27 ± 2.58	50.73 ± 2.53	67.12 ± 2.16	64.14 ± 2.22
AVES-Bio + OW100% of data	51.54 ± 2.54	53.01 ± 2.63	70.14 ± 2.13	68.49 ± 2.08
AVES-Bio + OW · BTB310% of data	57.86 ± 2.79	56.05 ± 2.48	78.08 ± 1.94	69.91 ± 2.06
AVES-Bio + OW · BTB350% of data	58.15 ± 2.60	56.98 ± 2.52	78.11 ± 1.92	71.19 ± 1.98
AVES-Bio + OW · BTB3100% of data	63.53 ± 2.93	58.91 ± 2.86	80.60 ± 1.84	73.93 ± 2.13

Task

Classification & detection after adaptation

Continue pretraining AVES-Bio on unlabeled OpenWhistle audio, then evaluate the adapted frozen encoder.

Classification: Macro-F1
Detection: mAP

Data-scale summary

80.80%Detection mAP

+4.04 ptsFull-set F1 vs 50% data

Checkpoint	Full set10 types · F1	Natural6 types · F1	Balanced6 types · F1	DetectionmAP
Wav2Vec2.0stride 960 · 10% · 40k steps	22.65 ± 1.07	22.11 ± 1.37	34.78 ± 2.18	20.76 ± 1.34
Wav2Vec2.0stride 960 · 50% · 200k steps	56.53 ± 2.61	59.34 ± 2.99	77.66 ± 1.98	77.03 ± 2.09
Wav2Vec2.0stride 960 · 100% · 400k steps	60.57 ± 2.98	61.22 ± 3.04	80.83 ± 1.87	80.80 ± 1.91

Task

Classification & detection across data scales

Pretrain Wav2Vec2.0 with stride 960 from scratch on 10%, 50% and 100% of OpenWhistle, then compare frozen encoders.

Classification: Macro-F1
Detection: mAP

Best end-to-end result

59.54%Full-set macro-F1

77.02%Detection mAP

Backbone	Full set10 types · F1	Natural6 types · F1	Balanced6 types · F1	DetectionmAP
AVES-Coreend-to-end	57.29 ± 1.73	61.24 ± 4.58	77.37 ± 0.85	74.10 ± 1.26
BioLingualend-to-end	54.32 ± 2.79	57.55 ± 1.46	78.99 ± 1.56	72.26 ± 0.44
AVES-Bioend-to-end	59.54 ± 1.05	62.06 ± 0.89	79.60 ± 0.62	77.02 ± 1.31

Task

End-to-end classification & detection

Optimize the encoder and prediction head jointly on expert-labeled OpenWhistle examples.

Classification: Macro-F1
Detection: mAP

Higher is better. Full set evaluates all ten expert labels; Natural keeps the observed frequencies of six types; Balanced gives those six types equal representation. Linear-probe uncertainty is estimated by bootstrap; fine-tuning reports variation across three seeds.

06 / WHAT COMES NEXT

A benchmark today.
A research substrate tomorrow.

Vocal development

Trace how a learned repertoire evolves across years and life events.

Social dynamics

Model exchanges, turn-taking and interactions inside contiguous sequences.

Label efficiency

Explore few-shot, active and semi-supervised learning where expert time is scarce.

Temporal shift

Measure robustness as pod composition and acoustic conditions change.

07 / TEAM

Built across disciplines.

Faadil Mustun¹ · Chiara Semenzin^1,2 · Roberto Dessì³ · Pablo Robin Guerrero¹
Pierre Orhan⁴ · Alexis Emanuelli¹ · Emanuele Rossi⁵
Yair Lakretz⁶ · Gonzalo de Polavieja⁷ · Germán Sumbre¹

¹ Institut de Biologie de l’ENS (IBENS), Département de biologie, École normale supérieure, CNRS, INSERM, Université PSL, Paris, France · ² Earth Species Project · ³ Not Diamond, San Francisco, USA · ⁴ Institut du Cerveau, Paris, France · ⁵ Sapienza University of Rome, Rome, Italy · ⁶ École Normale Supérieure, Paris, France · ⁷ Champalimaud Foundation, Lisbon, Portugal

CITATION

Cite our work.

If you use OpenWhistle in your research, please cite our paper.

@article{mustun2026openwhistle,
  title={OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations},
  author={Mustun, Faadil and Semenzin, Chiara and Dessì, Roberto and Robin Guerrero, Pablo and Orhan, Pierre and Emanuelli, Alexis and Rossi, Emanuele and Lakretz, Yair and de Polavieja, Gonzalo G. and Sumbre, Germán},
  journal={arXiv preprint arXiv:2609.34839},
  year={2026},
  url={https://arxiv.org/abs/2609.34839}
}

OpenWhistle

Bioacoustic datasets go broad.OpenWhistle goes deep.

whistles

hours of audio

years

expert labels

Every contourhas a character.

One pipeline.Two datasets.

In-domain listeningmakes the difference.

Linear-probe classification & detection

Classification & detection after adaptation

Classification & detection across data scales

End-to-end classification & detection

A benchmark today.A research substrate tomorrow.

Vocal development

Social dynamics

Label efficiency

Temporal shift

Built across disciplines.

Cite our work.

Bioacoustic datasets go broad.
OpenWhistle goes deep.

Every contour
has a character.

One pipeline.
Two datasets.

In-domain listening
makes the difference.

A benchmark today.
A research substrate tomorrow.