AI Decodes Sperm Whale Clicks: GAN Latent Variables Reveal 22 Coda Types' Structure

maier_ak · x · 2026-09-07

A 2026 paper by Beguš, Leban and Gero in Royal Society Open Science tackles how to study sperm whale communication without a Rosetta stone. Whales emit short click sequences called codas—roughly 22 catalogued types, with the two most common patterns accounting for two-thirds of recordings—yet their meaning remains unknown.

The key question: after training a deep generative model on raw whale-click audio, can we identify which acoustic features the model considers "important" for producing realistic codas, and thereby generate hypotheses about what whales encode? Since there's no ground-truth label for whale meaning, the team devised CDEV (Causal Disentanglement with Extreme Values), combining latent-space manipulation with causal-inference statistics to interrogate the model itself.

They trained a fiwGAN on 2,209 sperm-whale recordings spanning five coda types, using a five-bit latent code pushed to extreme values [+12.5]. CDEV showed latent bit 1 increases click count and timing regularity; bits 0 and 3 uncovered spectral mean and cross-click acoustic regularity. Setting those bits high reproduced a nine-click 9R coda. Running CDEV on human speech recovered known phonetic axes, validating the method despite using only a fraction of the whales' repertoire.

Related event: AI decodes sperm whale communication via latent-space interventions(2 posts)→

Original post →

More from Research

Research channel →