Project CETI: sperm whale codas
Project CETI's 2024 paper found four sperm whale coda features that combine freely; 2025-26 papers add vowel-like patterns. What codas mean is still unknown.
Open in the interactive tree →Sperm whales communicate in short click patterns called codas. Machine-learning teams analyse the Dominica Sperm Whale Project archive for structure, while the project's own plan puts tests with played-back sounds last.
As of October 2026
Sharma et al. (Nature Communications, May 2024) analysed 8,719 manually annotated codas of one Caribbean clan (at least 60 whales, 2005-2018) and described rhythm, tempo, rubato and ornamentation as freely combining features; the paper says the function of many codas is open, and its authors call the feature set a 'phonetic alphabet'. Beguš et al. reported vowel-like patterns (Open Mind, 2025) and, on 15 April 2026 in Proceedings of the Royal Society B, five properties with parallels in human phonology. WhAM, a model fine-tuned on 10,000 coda recordings to generate synthetic codas, was accepted at NeurIPS 2025 (preprint 1 December 2025). No playback results with sperm whales were found.
Open steps
- Test codas by playback Low AI leveragePlay natural or synthetic codas to sperm whales and record their responses; CETI lists playback validation as its last phase and no results were found.
- Grow the annotated coda archive Medium AI leverageThe 2024 analysis used 8,719 manually annotated codas from one clan; automatic click detection and annotation would let models learn from far more recordings.
- Check the vowel-like patterns Medium AI leverageTest whether the a- and i-coda patterns reported in 2025-26 hold in other clans and years, and whether whales respond differently to them.
- Link codas to behaviour Medium AI leverageMatch coda sequences with tag, drone and video records of what whales are doing, to find situations in which particular codas occur.
Where AI could help
Medium AI leverage. Models find structure in coda recordings, but meaning needs played-back sounds and behaviour data from whales at sea.
- Cluster codas and test which acoustic features carry information
- Generate synthetic codas as stimuli for later playback tests
- Match coda patterns with behaviour logged by tags and drones
Shown so far
- Sharma et al. (Nature Communications, May 2024) report that rhythm, tempo, rubato and ornamentation combine to explain variation across 8,719 codas, while the function of many codas stays open. source
- WhAM (NeurIPS 2025) is a transformer model fine-tuned on 10,000 coda recordings that generates synthetic codas, evaluated with Fréchet Audio Distance and perceptual studies with expert marine biologists (preprint, 1 December 2025). source