Is a dog’s bark simply an expression of emotion, or does it contain specific information that humans cannot understand? Artificial intelligence (AI) can already analyse thousands of animal sounds and identify patterns that would go unnoticed by the human ear. But researchers are asking a far more ambitious question: could technology one day help us understand what animals are actually saying?

According to Professor Rytis Maskeliūnas of the Faculty of Informatics at Kaunas University of Technology (KTU), this is already an established area of research, although we are still a long way from truly translating animal communication. Today, AI makes it possible to analyse vast amounts of audio data that would be impossible for a human to process and to identify recurring signals and their links with animal behaviour and the surrounding environment.

This opens the door to much more complex questions – not only what sound an animal makes, but also when it makes it, who it is directed at and under what circumstances.

“However, determining what information an animal is communicating from sound alone is far from straightforward. Acoustic analysis needs to be combined with behavioural observations and controlled experiments so that we can test whether an identified pattern genuinely has a specific communicative function. That is why I would place less emphasis on the idea of an imminent animal-language translator and more on the possibility of systematically uncovering, for the first time, what information is contained in animal communication,” says Maskeliūnas.

What Can a Dog’s Bark Tell Us?

Maskeliūnas and his colleagues began investigating the information contained in animal sounds several years ago. The researchers wanted to determine how much could be learned about a dog’s emotional state from its sounds alone.

For the study, the researchers collected recordings of sounds made by German Shepherds and, based on assessments by animal behaviour experts, analysed four categories associated with the dogs – happiness, anger, crying and loneliness.

“We did not analyse the dogs’ sounds simply as audio recordings. We broke them down into different frequency and time components and extracted acoustic features from them. We then used machine-learning methods to test whether these features could distinguish between different emotional states. The results showed that a considerable amount of information can be obtained from a dog’s voice – we were able to identify angry barking with fairly high reliability,” explains Maskeliūnas.

According to Maskeliūnas, however, the most important conclusion of the study was not the ability to determine whether a dog was angry or lonely. Rather, it showed that an animal’s voice contains structured information that the human ear cannot always detect.

“In other words, sound is not simply noise or a background expression of emotion – it contains acoustic features associated with the animal’s state. At the same time, this also highlights an important limitation: recognising a particular emotional state does not mean that we understand the entire signal or know exactly what information the dog is trying to communicate,” he emphasises.

According to Maskeliūnas, a similar study conducted today could be far more extensive. Instead of defining several emotional categories in advance and testing whether AI can recognise them, researchers could allow a model to independently analyse thousands or even millions of dog sounds and search for recurring structures within them.

“Specialised animal acoustic models are already being developed for this purpose. For example, Dog2vec was trained on more than 6,000 hours of recordings of dog barks and outperformed general-purpose models in some dog sound recognition tasks,” says Maskeliūnas.

Such technologies make it possible to analyse not only individual sounds but also their sequences, repetition and interaction with other signals, while linking this information with visual, behavioural and physiological data. Maskeliūnas notes that these multimodal approaches are already proving more effective in recognising dogs’ emotional states.

“The biggest breakthrough would not be for AI to say that the dog is angry, but for it to independently discover previously unknown combinations of signals and identify how they relate to particular situations and behaviours. The crucial question, however, is how we can reliably determine whether a pattern identified by AI is genuinely part of animal communication rather than simply a statistical feature of the data,” he says.

Decoding Animal Signals Requires More Than AI

One of the greatest challenges in studying animal communication, according to Maskeliūnas, is that recognising a signal is much easier than understanding what it means. AI may determine that a particular sound repeatedly occurs under certain circumstances – for example, when there is danger or when an animal greets its owner. But this pattern alone is not enough to establish what information the animal is conveying.

“Context plays an enormous role in animal communication: who is sending the signal, who it is directed at, what is happening around them at the time and how other animals respond. AI can identify patterns that humans have not previously noticed and generate a hypothesis about what a signal might mean, but statistical associations alone are not enough to confirm that hypothesis reliably – observations and experiments are still needed,” explains Maskeliūnas.

Researchers could be much more confident about the meaning of a particular signal, he adds, if changing the situation or presenting the same signal to another individual consistently produced a response that matched the AI model’s prediction.

“In this case, AI is a little like Sherlock Holmes – it spots an important clue, but the biological experiment is the forensic laboratory where we still have to establish whether that clue has been interpreted correctly,” says Maskeliūnas.

Another important capability of AI is its ability to analyse not only the sounds produced by an animal, but also visual information, its movements, the reactions of other individuals and environmental conditions at the same time. According to Maskeliūnas, this can reveal connections that humans themselves might overlook.

Such analysis makes it possible to move beyond the question of what kind of sound is being made and towards much more important ones: what happens before and after the animal produces the signal, and how other animals respond to it. Examining these relationships could bring researchers closer to understanding what a signal actually means.

“Research by other scientists shows that decoding animal communication must be based not only on similarities between acoustic signals, but also on how those signals are perceived by the receiving animal and how it responds to them,” explains Maskeliūnas.

Why Animal Communication Is Harder to Decode Than Human Language

Maskeliūnas points out that, from a technological perspective, some aspects of animal communication can be compared with human language, but this analogy should not be taken too literally. Large language models learn from enormous volumes of text in which a complex symbolic system already exists – words, grammar, syntax and meanings that humans understand.

“In animal communication, we usually do not have such a dictionary, and we do not know in advance whether an equivalent symbolic structure exists at all. This means we cannot simply teach AI an animal language in the same way that we train a human language model,” he explains.

Nevertheless, the technological principle remains broadly similar: vast amounts of data need to be analysed, recurring structures identified and what happens next predicted.

“The key difference is that, with human language models, we can tell the system what a particular word or sentence means. With animals, we do not have that map – we still need to discover it and verify it experimentally. First, we need to establish whether there is a structured communication system behind animal signals that it would actually make sense to call a language,” says Maskeliūnas.

According to Maskeliūnas, the most promising species for this kind of research are those with diverse, socially significant and sufficiently well-documented communication. These include whales and dolphins, elephants, certain primates and some bird species. Parrots, for example, can learn to imitate a variety of sounds, while crows are capable of making fairly complex decisions.

“Complex communication alone is not enough, however. AI needs not only large quantities of sounds, but also high-quality data that tell us what is happening to the animal at that moment, who the signal is intended for, how others respond to it and whether the same patterns recur across different individuals and situations. We simply do not yet have enough data of this kind,” explains Maskeliūnas.

The quantity of data also matters, he adds. Even a very capable model would struggle to identify reliable structures from only a few hours of recordings. At the same time, millions of recordings would not solve the problem on their own if their context were unknown.

“That is why the most promising species would not necessarily be the most intelligent or the most vocal, but those for which we can collect large, long-term and well-contextualised datasets. In this field, the quality of the data matters more than quantity alone,” emphasises Maskeliūnas.

Could AI Change the Way We Relate to Animals?

Maskeliūnas notes that AI-generated animal signals could help researchers test whether they genuinely understand their function. After analysing a large amount of data, for example, AI might determine that a particular signal is associated with a specific social situation. Researchers could then create a synthetic version of that signal and present it to an animal under controlled conditions.

If the animal consistently responded to the synthetic signal in the same way as it did to the natural one, there would be stronger grounds for believing that at least part of its function had been understood.

“In essence, this would be an experimental conversation – not in the sense that we had already learnt to talk to an animal, but in the sense that we could experimentally test our hypotheses. However, caution is essential here: a model may generate an acoustically convincing signal that does not carry the same meaning for the animal as the natural one. We would then have to determine whether the animal actually recognises and interprets the signal, rather than simply responding to its loudness, frequency or some other acoustic property,” he explains.

According to Maskeliūnas, if AI eventually allowed us to reliably understand even part of the information animals communicate, it could fundamentally change the relationship between humans and other species. Pet owners, for example, might be able to better understand their animals’ stress, fear, discomfort or needs, and identify problems earlier than would be possible from behaviour alone.

“In animal welfare, this could help us assess more objectively how animals feel under different housing or transport conditions, while in the wild it could allow us to monitor the condition of populations without interfering with their lives,” says Maskeliūnas.

Yet, according to him, the most significant change might have less to do with the practical applications of AI and more with the way humans perceive other species. If animal signals were found to have a far more complex structure than previously thought, conveying information about the environment, social relationships, danger or even individual experience, we might have to reconsider what we mean by intelligence and complex communication.
It would also raise moral questions – the more we understand about animals’ states and needs, the harder they become to ignore.