Sperm Whales Have an Alphabet: How AI Is Decoding Their Deep-Sea Conversations

For decades, marine biologists have recorded the powerful clicks of sperm whales and wondered whether these sounds form a kind of language. Unlike humpback whales, sperm whales do not produce long melodic songs. Instead, they emit rapid sequences of broadband clicks known as codas. These codas are used in social settings, especially among family groups, and they can travel long distances through deep ocean water.

Recent advances in artificial intelligence have allowed scientists to analyse these clicks at a scale never before possible. The emerging evidence suggests that sperm whale codas are built from structured acoustic elements that function in ways comparable to phonetic building blocks in human speech.

What Are Codas?

Codas are patterned sequences of clicks, usually consisting of three to forty individual pulses. Researchers first documented recurring coda types in the 1980s and 1990s. They observed that different whale groups, sometimes called clans, use distinct sets of codas, suggesting that the patterns are socially learned rather than random.

Long-term research conducted by the Dominica Sperm Whale Project has recorded thousands of these vocal sequences. By comparing click timing and spacing, scientists discovered that codas are not simple repetitions. They vary in rhythm, duration, and internal structure. This variation raised the possibility that codas carry information about identity, group affiliation, or behavioral context.

How AI Changed the Analysis

The breakthrough came when researchers began applying machine learning to massive acoustic datasets. Scientists working with MIT Computer Science and Artificial Intelligence Laboratory and Project CETI used advanced pattern recognition algorithms to examine more than 8,700 codas collected in the Caribbean. A study published in Nature Communications identified four main acoustic features that structure sperm whale codas: rhythm, tempo, rubato, and ornamentation. Rhythm refers to the spacing between clicks. Tempo describes the overall speed of the sequence. Rubato captures subtle variations in timing. Ornamentation involves the addition of extra clicks within a sequence.

By analyzing these features computationally, the AI models revealed that codas are not isolated signals but combinations of repeatable elements. In human speech, phonemes combine to form words. In sperm whales, combinations of acoustic features appear to generate a flexible repertoire of vocal patterns. Pratyusha Sharma, a lead researcher involved in the study, explained that this structured variability expands the potential expressive range of the whales’ communication system, even though researchers do not yet understand the meaning of specific sequences.

Pattern Recognition at Oceanic Scale

The scale of acoustic data collected from whales makes human-only analysis impractical. Machine learning systems excel at detecting subtle patterns across large datasets. Project CETI researchers developed neural network models that cluster codas according to acoustic similarity and context.

One experimental model, described in a 2025 preprint, is called the Whale Acoustics Model, or WhAM. This transformer-based neural network was trained on thousands of recorded codas and could generate synthetic sequences that preserved the statistical structure of natural whale clicks. Although the model does not translate whale speech, it provides a tool for testing hypotheses about how codas are constructed. This computational approach mirrors methods used in human linguistics and speech recognition, where AI identifies recurring units and hierarchical structure before researchers attempt semantic interpretation.

Signs of Combinatorial Complexity

One of the most intriguing findings is evidence of combinatorial organization. In human language, combining a limited number of sounds can create an enormous number of meaningful words. Similarly, sperm whale codas appear to combine basic acoustic features into larger structured sequences.

Researchers have found that whales modify click timing and grouping in response to social context, such as group coordination or caregiving interactions. These variations suggest that codas are not fixed signals but adaptable expressions. Linguists collaborating with Project CETI have noted parallels between certain whale acoustic patterns and phonetic systems in human language, though they caution that structural similarity does not prove shared meaning or grammar.

Caution and Scientific Restraint

Experts emphasize that identifying structured sound patterns does not mean scientists have decoded whale language. Jacob Andreas, a computational linguist at MIT involved in the research, has stated that understanding structure is only the first step. Linking specific codas to behavioral outcomes is essential before claims about meaning can be made.

David Gruber, founder of Project CETI, has also stressed that the goal is not to anthropomorphize whales but to rigorously investigate the possibility that they possess complex communication systems adapted to deep-sea life. At present, researchers can describe how codas are organized, but not what they signify. The distinction between acoustic structure and semantic content remains central to responsible interpretation.

A New Frontier in Marine Science

The application of AI to sperm whale vocalizations represents a turning point in the study of animal communication. By revealing a structured, combinatorial system built from repeatable acoustic elements, scientists have shown that sperm whale communication is more complex than previously assumed. Whether this system ultimately qualifies as language in the human sense remains uncertain. What is clear is that advanced computational tools are enabling researchers to map the architecture of whale vocal systems with unprecedented precision.

In the vast darkness of the deep ocean, sperm whales exchange sequences of clicks that may form a rich and layered communication network. Artificial intelligence is now helping scientists begin to understand that network, one patterned click at a time.