MATH · IN · MODELS

Dolph2Vec (Wav2Vec2-style SSL model for dolphin vocalizations)

ENS Paris / Institut du Cerveau / Champalimaud Foundation

Structures found in this family (1)

Linear Separability

By model (1)

Dolph2Vec (Wav2Vec2.0 architecture adapted for 44.1kHz dolphin audio, trained on 100hrs/5yr real longitudinal recordings)
Linear Separability

Observations (1)

A real Wav2Vec2-style model trained on longitudinal dolphin recordings organizes its codebook by whistle category and individual identity

Papers

Dolph2Vec: Self-Supervised Representations of Dolphin Vocalizations (2026)