How to speak whale

Each day, a sperm whale needs to eat 3 to 4% of its body weight to survive – and for a species that can weigh up to 57 tons, this means a staggering quantity of food. By some estimates, a fully grown sperm whale can eat more than a ton of food in 24 hours.

Depth and dive shape are illustrative

Dominica’s year-round community offers the chance to observe exactly how they manage this. Project CETI uses boats to gather unique video and has developed a method for attaching complex sensors to the whales’ gigantic bodies. They fly quadcopter drones that stick advanced wildlife tracking devices directly on to the side of the whale. These biologger devices attach with non-harmful suckers and contain GPS trackers with depth sensors, movement trackers, heart rate monitors and more.

The resident whales embark on deep “foraging dives”, spending 45 minutes to an hour, descending from the warm sunlit waters around Dominica into the extreme pressure and near-complete darkness of around 700 metres depth, explained Petri.

“When they resurface, they spend about 15, 20 minutes breathing, re-oxygenating after long dives,” he said. This is the basic routine played out, again and again, with some whales synchronizing their deep dives.

In Dominica, Project Ceti employs underwater microphones known as hydrophone, and is developing semi-autonomous marine “gliders” which could follow whales for weeks without a human driver. These together can detect sperm whales’ vocalizations. They show that they emit a rhythmic series of short clicks that researchers call codas. Researchers believe these codas are “like little words,” said Petri.

Rings follow a real recorded two-whale exchange

“Most of the socialisation that we know of happens on the surface,” explained Petri. “So when they come up, often what you see is that animals that are a few hundred metres apart start drifting toward each other, and then they start coda-producing rapidly once they are very close, within one body length of each other. Then, they start chatting a lot,” he said.

These regular interactions give rise to a set of questions that are driving current research: are they coordinating their synchronised dives? Are some whales leading and others following? And what other information might they be passing on?

In Dominica, all of this data is streamed – from biologgers, hydrophones, and gliders – into a massive “data vault,” explained Petri. Then comes one of the hardest bits – annotating the acoustic data so that it can be analyzed by data scientists.

First, the system will detect clicks and identify whether they are the kinds of regular pinging clicks sperm whales use in the dark, deep ocean for echolocation, or the rhythmic clicks they emit when socialising.

“Then you want to sort of figure out which clicks go together in one coda,” Petri said. This can be like trying to listen into a busy party, he said “because you might have multiple animals speaking at the same time, or you might have noise, so it might be that some clicks are missing, and so you have to infer.”

A process of source-separation sorts which of the resident whales – which the team have named – is talking.

“So you want to say, well, this coda comes from Pinchy,” one of the named whales, he said.”This other coda is actually Digit talking.”

Ultimately this aims to translate sound into usable data. A large part of this is automated using AI pattern-recognition systems, and the team is working on making it more effective, “but we also still have humans in the loop”, including the Project CETI biology team who are trained to help curate and check that the automatic annotation actually makes sense,” Petri said.

“Now, once you have the clean annotated acoustic data, then you can pair it with behavioural data, for example, the accelerometers, the depth sensors, and so on and on,” said Petri.

This has been a driving ambition of Project CETI’s work in Dominica, to develop sustained monitoring that can pair sound and action. Together, they hope these can provide something like a Rosetta Stone for whale language, showing which repeated codas correlate with key events.

Already, simply by feeding recordings of whale chat into machine learning models, the Project CETI team has found that around eight to ten codas can predict whether they are about to head off on deep foraging dives or not, explains Pertri.

But this is far from cracking whale language. The CETI team aims to understand the basic units of language. The current working hypothesis is that they have a basic set of symbols – a sort of fundamental phonetic alphabet – that have been identified using neural networks, which extract measurable features from the sound. By examining the timing, rhythm, spectral shape of codas, they have identified two vowel-like sounds and recent research has shown that they may modulate these, like tones in Mandarin, to express different senses. But their meaning is not yet known, Petri said.

Petri is a theoretical physicist who wrote his thesis at CERN, the European Organization for Nuclear Research, where scientists explore the fundamental structure of the Universe. But he was invited to join Project CETI at the outset, as an expert in topological data analysis, which looks at data points as a connected network and uses mathematical shapes to find hidden patterns in complex information.

His team uses statistical, network and machine-learning tools to identify recurring patterns and relationships, which can help ultimately shed light on the meaning of the codas.

“AI is fundamental for us, from coordinating flying fleets of drones to finding what the regularities are in the language that they have,” said Petri.

In a previous paper in the journal Science, Petri used the same kinds of network science tools that his colleagues at Northeastern NetSI institutes use to trace how viral infections spread through hospitals or misinformation spreads through social interaction. This showed for the first time how two different family units cooperated to help a newborn calf to breathe in its first hours after birth, while it is still becoming buoyant and learning to swim. This analysis allowed them to see which whales each individual is paying attention to, with attention shifting from the mother before the birth, to the calf post-partum.

To understand more about foraging dives, they aim to use these systems to find out if certain influential whales are initiating deep dives, and others are following.

“Is there a clear leadership within this community – are older females typically the ones beginning dives or not? Are the calves the ones that are driving communication by basically going around and chatting with everyone?” Petri said.

From here, it should become clear whether there is a particular coda – a whale word – that triggers a dive.

Petri recalls two whales diving together in perfect sync, with biologgers attached, allowing the team to record “both acoustically and behaviorally” how this coordination took place. Over time, another whale, named Peach, drifted closer and joined the yo-yoing process of feeding and resting, explained Petri.

“So for the last few dives, there are actually the three of them diving together going up and down in very close proximity,” he said.

Petri’s team is using proximity networks and acoustic analysis to investigate how it all came together. They aim to understand if echolocation plays a role.

As sperm whales dive into the ink black depths, their eyesight fails and they use clicks to sense squid. In the dark, Project CETI has hypothesized they may eavesdrop on each other’s clicks, to sense where large groups of squid are – like two people at night in a forest who are “using the reflection of each other’s torch lights to find their way in the dark,” said Petri.

If whales do use one another’s clicks to find prey, the sounds could serve simultaneously as both communication and perception.