AI Discovers an Alphabet in Whale Calls

Researchers have revealed a complex communication structure in sperm whales by analyzing their click sounds with artificial intelligence. The discovery of a "phonetic alphabet" in their calls could be a step toward understanding whales' social behaviors and messages.

Share
AI Discovers an Alphabet in Whale Calls

Researchers have revealed a new level of complexity in whale communication by analyzing thousands of recorded sequences of sperm whale click sounds using artificial intelligence.

The Whales' Soundscape

Variations in tempo, rhythm, and length of the whales' click sequences, called codas, create a rich acoustic tapestry. These variables suggest whales can combine click patterns in multiple ways, mixing and matching phrases to convey a wide range of information to each other.

What sperm whales are saying with their clicks remains a mystery to human ears. But revealing the extent of whales' vocal exchanges is an important step toward linking whale calls to specific messages or social behaviors, the researchers reported on May 7 in the journal Nature Communications.

A Deeper Understanding of Whale Codas

"This work builds on a lot of prior research focused on understanding sperm whale calls. However, this is the first time researchers have begun studying their calls in a broader communicative context and in exchanges between whales, which enabled some of these findings," said study co-author Dr. Daniela Rus, director of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT.

Understanding which aspects of their codas they can control and vary helps us understand how they might encode information in their calls, Rus explained.

The researchers compared their catalog of sound combinations to a "phonetic alphabet" for sperm whales, in which variations in the whales' click sequences were likened to the production of different phonetic sounds in human speech.

Skepticism Toward the "Phonetic Alphabet"

Despite the interesting findings, Dr. Luke Rendell, a researcher at the University of St. Andrews in the UK, believes the term "phonetic alphabet" may be misleading when it comes to whales' vocal interactions.

"The way tempo variation is used is completely different from how we use elements of an alphabet to construct linguistic expressions," Rendell said. "There's no evidence for this, and it's not a particularly helpful interpretation, since it forces everything into a narrow, somewhat oversold view of whether it resembles human language or not, when there's a much broader range of interpretations available."

Discovering Patterns With AI

Sperm whales produce their clicks by forcing air through an organ in the head called the spermaceti, and these sounds can reach as high as 230 decibels — louder than a rocket launch and capable of rupturing human eardrums, another research team reported earlier in the journal Scientific Reports.

For the new study, the researchers used machine learning to detect patterns in sound data collected by The Dominica Sperm Whale Project. Machine learning, a branch of AI, enabled the analysis of large amounts of complex sound data, which would have been extremely time-consuming and difficult for humans to do manually. The recordings represented the voices of about 60 sperm whales, part of a group of about 400 whales known as the Eastern Caribbean clan, and the vocalizations were recorded between 2005 and 2018.

The researchers examined the timing and frequency of 8,719 coda sequences — in individual whales' utterances, in choruses, and in call-and-response exchanges between whales. When visualized with artificial intelligence, previously unseen coda patterns emerged.

Complexity in Whale Communication

The study's authors defined four properties of codas: rhythm, tempo, rubato, and ornamentation. Rhythm describes the sequence of intervals between clicks. Tempo is the length of the entire coda. Rubato refers to variations in length across adjacent codas with the same rhythm and tempo. And ornamentation is an "extra click" added to the end of a coda in a group of shorter codas, Rus explained.

These so-called ornamental clicks "occur more toward the beginning and end of turns" during vocal exchanges between whales, "and behave like discourse markers," Rus said.

The discovery that whales could synchronize variations in coda tempo was "a very interesting observation," Rendell said.

"I'm less convinced about the 'ornamentation,'" he added. "It occurs very rarely, and I think we need more evidence that they aren't just production glitches," or filler sounds, "like when we say 'um' or 'err.'"

Potential Applications

The program identified a total of 18 types of rhythm, five types of tempo, three types of rubato, and two types of ornamentation. These coda features could all be mixed and matched to create a "huge repertoire" of phrases, the study's authors reported. In addition, meaning could be further adjusted depending on the placement of a coda — following or overlapping other codas — within an exchange or chorus involving two or more whales.

Interactive experimentation with whales, together with observations of whale behavior, could be an important part of decoding the syntax of sperm whale click sequences, the authors wrote in the study. Their machine learning method could also prove useful for studying other types of animal vocalizations, added Dr. Brenda McCowan, a professor at the University of California, Davis School of Veterinary Medicine.

"Tempo, rhythm, rubato, and ornamentation are likely to be found in other whale species," McCowan said. "We already know this is true for humpback whale song. But there's also evidence of this type of pattern in other aquatic, terrestrial, and arboreal species to which this method could be applied."

But even though this technique is helpful for identifying certain aspects of communication, it's no Rosetta Stone, Rendell cautioned.

"Machine learning is good at finding patterns in large datasets," he said, "but it doesn't create meaning."