Chiara Semenzin

Mustun, F.*, Semenzin, C.*, Dessì, R., Robin Guerrero, P., Orhan, P., Emanuelli, A., Rossi, E., Lakretz, Y., de Polavieja, G., Sumbre, G. (2026). OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations. NeurIPS 2026 Datasets & Benchmarks Track. Spotlight [arXiv] [Website]

The largest publicly available dolphin vocalization dataset: ~180,000 whistles (114 hours) recorded over five years from a stable pod of five bottlenose dolphins, with 8,354 expert-annotated whistles and evaluation protocols for whistle-type detection and classification. A Wav2Vec2.0 model pretrained on the corpus outperforms large foundation bioacoustics models on both tasks.

Semenzin, C., Mustun, F., Dessì, R., Emanuelli, A., Orhan, P., Lakretz, Y., de Polavieja, G., Sumbre, G. (2025). Dolph2Vec: Self-Supervised Representations of Dolphin Vocalizations. NeurIPS 2025 Workshop on AI for Non-Human Communication. [arXiv]

Dolph2Vec, self-supervised, species-specific model significantly outperforms general-purpose baselines in detection and classification. The learned embeddings capture interpretable dolphin whistle categories, enabling fine-grained analysis of communication patterns.

Mustun, F.*, Semenzin, C.*, Rance, D., Marachlian, E., Guillerm, Z., Mancini, A., Bouaziz, I., Fleck, E., Shashar, N., de Polavieja, G., Sumbre, G. (2024). Whistle variability and social acoustic interactions in bottlenose dolphins. In Review

Signature whistles vary systematically, forming distinct sub-categories that mirror social structures and serve different communicative roles. Remarkably, dolphins were also observed producing the signature whistles of their deceased mothers—a behavior previously seen only in humans.

Semenzin, C., Hamrick, L., Seidl, A., Kelleher, B., Cristia, A. (2021). Towards large-scale data annotation of audio from wearables: validating zooniverse annotations of infant vocalization types. IEEE Spoken Language Technology Workshop (SLT), pp. 1079–1085.

Classification of individual vocalizations on Zooniverse was overall moderately accurate compared to the laboratory gold standard.

Semenzin, C., Hamrick, L., Seidl, A., Kelleher, B.L., Cristia, A. (2021). Describing vocalizations in young children: A big data approach through citizen science annotation. Journal of Speech, Language, and Hearing Research.

A large dataset of infant speech was uploaded on a citizen science platform. The same data were annotated in the laboratory by highly trained annotators. An analysis of descriptors defined at the level of individuals found strong correlations between descriptors derived from Zooniverse versus laboratory annotations.