AI as a tool for discovery in animal communication
summary
Claims that sperm whale coda vowels exist and behave like human vowels, supported by a GAN-based artificial baby language learner that predicts spectral structure later confirmed with spectrograms and LPC. Key specifics: cutting inter-click silences reveals formant-like bands across common codas (e.g. 1+1+3, 5R) and longer creaks (18–28 clicks); first-click prominence, diphthongs and coarticulation recur; about 10–15% of codas are ambiguous. AI interpretability (single-unit interventions) pinpoints where to look, but manual inspection and underwater tag data validate the findings.
AI-written summary · about
chapters (13)
- 0:00AI in animal communication
- 1:10AI as discovery tool
- 3:38Ecologically plausible language models
- 10:35AI interpretability in whale calls
- 14:44Human vowels spectral properties
- 17:58Coda vowels in sperm whales
- 22:49Source filter theory in whales
- 27:54Coda vowels like human vowels
- 35:53Confirming AI predictions
- 37:00Human role in AI discovery
- 37:41Three intelligences informing each other
- 48:10Classifying coda vowels
- 50:11Q&A
related talks
01h 03mGenerative AI and What Is Meaningful Sperm Whale Communication
53mWhAM: Towards A Translative Model of Sperm Whale Vocalization
01h 04mGPT-Whale Language Models at Tools for Structure Discovery and Hypothesis Generation
01h 00mLong term work and Machine Learning Approaches to DolphinGemma
25mThe Sound of Life
01h 07mDecoding Dolphin Speech
57mDecoding Dolphin Communication
42mArtificial Intelligence That Removes Language Barriers