The race to translate animal sounds into human language

In 2025 we will see artificial intelligence and machine learning being harnessed to make real progress in understanding animal communication, answering a question that has baffled humans for as long as we have existed: “What do animals say to each other?” the recent Coller-Dolittle Awardoffering cash prizes of up to half a million dollars to scientists who “crack the code” is an indication of bullish confidence that recent technological developments in machine learning and large language models (LLMs) are putting this goal within reach .

Many research groups have been working for years on algorithms to make sense of animal sounds. The Ceti Project, for example, has been decoding the click trains of sperm whales and the songs of humpback whales. These modern machine learning tools require extremely large amounts of data, and until now, such amounts of high-quality, well-annotated data have been lacking.

Consider LLMs like ChatGPT that have training data available that includes the entirety of the text available on the Internet. This information about animal communication has not been accessible in the past. It’s not just that human data corpora are many orders of magnitude larger than the kind of data we have access to on animals in the wild: over 500GB of words were used to train GPT-3, compared with just over 8,000 “codas”. ”(or vocalizations) for the recent Ceti Project analysis of sperm whale communication.

Furthermore, when we work with human language, we already know what is being said. We even know what constitutes a “word,” which is a huge advantage over interpreting animal communication, where scientists rarely know whether a particular wolf howl, for example, means anything different from another wolf howl, or even if wolves consider a howl to be somehow analogous to a “word” in human language.

However, 2025 will bring new advances, both in the amount of animal communication data available to scientists and in the types and power of AI algorithms that can be applied to that data. Automated recording of animal sounds has become available to all scientific research groups, and low-cost recording devices such as AudioMoth are gaining popularity.

Massive data sets are now coming online, as recorders can be left in the field, listening to the songs of gibbons in the jungle or birds in the forest, 24/7 , for long periods of time. There were times when it was impossible to manually manage such massive data sets. Now, new automatic detection algorithms based on convolutional neural networks can go through thousands of hours of recordings, selecting animal sounds and grouping them into different types, based on their natural acoustic characteristics.

Once such large animal data sets become available, new analytical algorithms become a possibility, such as using deep neural networks to find hidden structures in sequences of animal vocalizations, which may be analogous to the meaningful structure of language. human.

However, the fundamental question that remains unclear is: what exactly are we hoping to do with these animal sounds? Some organizations, like Interspecies.io, state their goal quite clearly: “translate signals from one species into coherent signals for another.” In other words, to translate animal communication to human language. However, most scientists agree that non-human animals do not have their own language, at least not in the same way that humans have language.

The Coller Dolittle Prize is a little more sophisticated and seeks a way to “communicate or decipher the communication of an organism.” Deciphering is a slightly less ambitious goal than translating, considering the possibility that animals do not, in fact, have a language that can be translated. Today we do not know how much or how little information animals transmit to each other. By 2025, humanity will have the potential to surpass our understanding not only of how much animals say but also what exactly they say to each other.