WhaleLM: Finding Structure and Information in Sperm Whale Vocalizations and Behavior with Machine Learning

Pratyusha Sharma, Shane Gero, Daniela Rus, Antonio Torralba, Jacob Andreas

TL;DR

Researchers have developed a "whale language model" using machine learning to decode the complex click sequences (codas) of sperm whales. This model reveals that whale vocalizations have a sophisticated structure, including long-range dependencies and predictable turn-taking, and can even predict future whale behavior with high accuracy.

What Problem Does This Paper Address?

Understanding animal communication is a significant scientific challenge. While we know many animals communicate, the intricate structure and meaning behind their signals remain largely unknown. Sperm whales, in particular, use complex sequences of clicks called codas during coordinated activities like foraging, but it's unclear if these sequences follow systematic patterns or convey specific information about their behavior. This paper aims to bridge this gap by applying advanced machine learning techniques, similar to those used for human language, to decipher the structure and informational content of sperm whale vocalizations.

Methods

The core of this research lies in applying neural sequence models, specifically Long Short-Term Memory (LSTM) networks, to sperm whale vocalizations. These models are trained to predict the next vocalization (coda) in a sequence, or to predict the whale's behavior based on its vocalizations. A key methodological innovation is the use of "ablation studies." By systematically removing certain features of the codas (like rhythm or tempo) or limiting the historical context the model can access, the researchers can quantify which aspects of the vocalizations are most informative for prediction. This allows them to generate hypotheses about the structure and function of whale communication without prior human interpretation.

Datasets

The study utilizes a dataset of sperm whale coda exchanges collected by The Dominica Sperm Whale Project (DSWP) between 2014 and 2018. This dataset includes bio-logging tag data (Dtags) from known individuals in the Eastern Caribbean clan (EC1), providing audio recordings of codas and synchronized accelerometer, gyroscope, and magnetometer data to track whale behavior (e.g., diving depth).

Key Insights and Ideas

  • Whale communication is more than just individual sounds: The research demonstrates that the order and sequence of codas are crucial for conveying information, not just the individual types of codas themselves.
  • Long-range dependencies exist: Sperm whale vocalizations show dependencies on codas produced up to eight steps prior (roughly 30 seconds), indicating a sophisticated, context-aware communication system.
  • Vocalizations predict behavior: The model can predict not only future whale vocalizations but also their current and future behavioral states (like diving) with remarkable accuracy, suggesting that codas play a role in coordinating group actions.
  • Machine learning as a hypothesis generator: The approach of using sequence models to identify informative features and patterns offers a powerful, data-driven way to generate hypotheses about animal communication, which can then be tested with traditional methods like playback experiments.

Key Results

  • Order matters: Removing the order of codas in the input data decreased prediction accuracy by up to 22.7%.
  • Context is key: Limiting the model's context window to fewer than six preceding codas reduced predictability by up to 20.6%.
  • Behavior prediction accuracy: The model achieved 72.8% accuracy in predicting the whale's current behavioral state and an impressive 86.4% accuracy in predicting future behavior, significantly outperforming chance baselines.
  • Rhythm and tempo are informative: While rhythm and tempo features are important for predicting immediate future calls, the rhythm feature becomes less critical with longer contexts, suggesting redundancy with other features over time.
  • Group signaling before dives: The study provides evidence that diving behavior is often preceded by multiple whales producing the same "prototypical" coda, suggesting a form of group signaling or consensus formation.

Limitations and Future Work

The authors acknowledge that their machine learning approach identifies associational relationships (similar to "Granger causality") between vocalizations and behaviors, but it does not definitively establish causal links. Future work could involve more expressive models and larger datasets to uncover more complex dependencies. The researchers also suggest that their framework can be extended to incorporate social context (identifying individual whales) and to understand the whales' own goals and environmental observations. Ultimately, this work provides a strong foundation for future experimental investigations into the meaning and function of sperm whale communication.

Summary updated

Abstract

Abstract Language models (LMs), which are neural sequence predictors trained to model distributions over natural language texts, have come to play a central role in human language technologies like machine translation and information retrieval. They have also contributed to the scientific study of human language itself, enabling progress on long-standing questions about the learnability, optimality, and universality of key features of human languages. Many analogous questions exist in the study of communication between non-human animals—for which, in many cases, we have only a preliminary understanding of signals’ structure and use. Can neural sequence models help us understand these animal communication systems as well? We use these models to characterize the structure and information content of sperm whale vocalizations. Sperm whales ( Physeter macrocephalus ) engage in complex, coordinated behaviours like foraging and navigation in the darkness of the ocean while exchanging sequences of rhythmic clicks known as codas. However, little is known about whether there are any systematic patterns governing coda production, or how codas influence group decision-making and behaviour. To begin to answer these questions, we first train a neural sequence model (a ‘sperm whale language model’) to predict whales’ future vocalizations from their conversational history. By systematically manipulating the information available to this model, and measuring the change in predictive accuracy, we show that sperm whale vocalizations exhibit order dependence, long-range dependencies on up to the past eight codas in an exchange, and predictable turn-taking. Second, we train the sequence model to predict whales’ behaviour from their vocal exchanges, and find that both current behavioural context and future actions are predictable, with accuracies of 72% and 86% respectively, from coda sequences. Our study provides the first evidence that sperm whale vocalizations contain information that could be used to coordinate behaviour. More generally, it offers a framework for using modern machine learning tools for hypothesis generation and to assist in investigating the structure and function of unknown communication systems.
BibTeX
@inproceedings{sharma2025whalelm,
  author = {Pratyusha Sharma and Shane Gero and Daniela Rus and Antonio Torralba and Jacob Andreas},
  title = {{WhaleLM: Finding Structure and Information in Sperm Whale Vocalizations and Behavior with Machine Learning}},
  booktitle = {AiForAnimalComms @ NeurIPS},
  year = {2025},
  url = {https://openreview.net/forum?id=ppCEqgYXHd},
}
Figure