Project CETI: Listening to Sperm Whales with Machine Learning

Project CETI (Cetacean Translation Initiative) is an independent non-profit research organization applying advanced machine learning, robotics, and non-invasive bioacoustic sensors to study the communication of sperm whales (Physeter macrocephalus). Founded in 2020 by marine biologist and National Geographic Explorer David Gruber alongside a multidisciplinary consortium of computer scientists, oceanographers, and roboticists, Project CETI centers its primary field operations on the resident sperm whale clans off the island of Dominica in the Eastern Caribbean.

While popular media frequently sensationalizes the initiative as an effort to build a "two-way whale translator," CETI's actual peer-reviewed work focuses on deciphering acoustic structure, context, and combinatorial exchange rules through rigorous empirical science.

Focal Species and Field Methodology in Dominica

Sperm whales are the largest toothed predators on Earth, possessing the largest brain of any known species. Rather than producing frequency-modulated whistles like dolphins, sperm whales communicate almost exclusively through patterns of broadband clicks called codas:

Individual Click:  ║ (duration < 1 ms; broadband pulse)
Click Coda:        ║  ║  ║   ║    ║ (e.g., "1+1+3" rhythm pattern; duration < 3 s)

The concept of codas was first established in 1977 by William A. Watkins and William E. Schevill at the Woods Hole Oceanographic Institution:

Project CETI Multi-Sensor Field Architecture

Studying sperm whales in deep oceanic waters requires overcoming severe observational constraints. CETI combines three synchronized sensory platforms:

1. Non-Invasive Suction-Cup Tags (DTAGs / CETI Tags): Deployed via long carbon-fiber poles or autonomous drones, soft suction cups attach tags to whale skin. On-board sensors record high-resolution hydrophone audio (sampled at up to 96 kHz), 3D triaxial accelerometry, depth pressure, and inertial orientation.
2. Synchronized Autonomous Hydrophone Arrays: Fixed vertical mooring arrays deployed in deep water record acoustic arrivals across multiple depth channels, allowing automated time-difference-of-arrival (TDOA) acoustic localization.
3. Aerial Drone Photogrammetry: High-resolution drone video tracks surface behaviors, spatial spacing, heading, and mother-calf nursing configurations synchronized to the millisecond with underwater hydrophone tracks.

The Dominica Sperm Whale Project Baseline

Project CETI's computational investigations build upon two decades of foundational field biology conducted by the Dominica Sperm Whale Project (DSWP), founded in 2005 by Dr. Shane Gero:

• Cultural Clans: Sympatric clans of sperm whales share distinct vocal dialects, as demonstrated by Rendell and Whitehead (2003):

• Multi-Level Identity Cues: Gero, Whitehead, and Rendell (2016) demonstrated that Eastern Caribbean codas contain nested acoustic markers that simultaneously broadcast the identity of the individual whale, its family unit, and its broader vocal clan:

Groundbreaking Finding: Combinatorial Coda Structure (Sharma et al. 2024)

In May 2024, researchers from MIT CSAIL and Project CETI published a landmark study in Nature Communications authored by Pratyusha Sharma, Shane Gero, Roger Payne, David F. Gruber, Daniela Rus, Antonio Torralba, and Jacob Andreas:

Analyzing a dataset of 8,719 codas recorded from the Eastern Caribbean clan between 2005 and 2018, the authors demonstrated that sperm whale codas are not rigid, isolated acoustic templates. Instead, codas exhibit continuous combinatorial variations along four structured acoustic axes:

[ Coda Acoustic Combinatorial Space (Sharma et al. 2024) ]
 ├── 1. Rhythm: The basic discrete pattern of inter-click intervals (e.g., 1+1+3).
 ├── 2. Tempo: Overall coda duration continuously modulated across conversational bouts.
 ├── 3. Rubato: Systematic acceleration or deceleration of click timing within a coda,
 │    matching the tempo shifts of interacting social partners.
 └── 4. Ornamentation: Addition of extra, non-canonical clicks at the start or end of a coda.

By showing that rhythm and tempo vary independently, and that whales systematically coordinate rubato shifts during vocal exchanges, the study demonstrated that the sperm whale communicative repertoire is combinatorial and expressive, rather than a small, static set of discrete signals.

Distinguishing Findings, Objectives, and Open Questions

To maintain scientific rigor, researchers must distinguish three distinct tiers of CETI's work:

1. Demonstrated Empirical Findings (Published):

   • The Eastern Caribbean sperm whale clan utilizes a combinatorial coda repertoire where tempo and rubato vary continuously alongside discrete rhythm types (Sharma et al., 2024).
   • Codas encode nested multi-level identity cues distinguishing clan, social unit, and individual caller (Gero et al., 2016).
   • Sympatric vocal clans maintain culturally transmitted acoustic dialects without genetic isolation (Rendell & Whitehead, 2003).

2. Project Engineering Objectives (In Progress):

   • Developing high-density suction-cup recording tags (CETI tags) with high-frequency hydrophones, depth pressure transducers, and 9-axis inertial measurement units.
   • Deploying deep-water autonomous mooring arrays to perform automated multi-channel time-difference-of-arrival (TDOA) acoustic localization.
   • Building deep learning models for automated click detection, biological vs. anthropogenic noise separation, and multi-whale cocktail party attribution.

3. Unresolved Scientific Questions (Open Frontiers):

   • Semantic Referentiality: Current evidence establishes statistical combinatorial structure and coordinated vocal turn-taking. It does not demonstrate that codas carry semantic meaning, referential nouns/verbs, or symbolic human-language translation equivalents. Whether coda variations denote external referents or convey social affiliation and arousal remains unknown.
   • Geographic Generalizability: Findings to date are derived from the resident Eastern Caribbean clan (~8,719 codas). Whether Pacific, Indian Ocean, or polar clans share this four-dimensional combinatorial system or employ simpler acoustic repertoires remains to be investigated.

For more information on the initiative's field expeditions and open-access data frameworks, visit the Project CETI official website.