How AI could help us talk to animals

Way back in the 80s, Joyce Poole, an elephant communication researcher, noticed a fascinating behavior among elephants. When one called out to her family, sometimes only one individual responded while others ignored the call. Intrigued, she speculated that elephants might be capable of directing calls to specific individuals, but lacked the means to prove it.

Decades later, Poole partnered with Mickey Pardo, who designed a study based on her observations. They recorded elephant calls in the field, carefully noting who made each call, who it was directed to, and the context. By encoding this acoustic information into numerical data, they fed nearly 500 different calls into a statistical model. Remarkably, the model could predict the intended recipient of a call better than chance, suggesting that African savanna elephants might have a naming system.

When they shared their findings on social media, one commenter noted, “the Earth just shifted a little bit,” capturing the significance of this discovery. This is just one example of how machine learning is unlocking complexities in animal communication that are beyond human detection. Some researchers are now looking to create large language models designed for interspecies communication, similar to the technology behind chatbots.

To study animal communication, researchers typically record vocalizations, observe behaviors, and conduct playback experiments. However, real-life recordings often resemble a “cocktail party,” with multiple animals vocalizing at once. This presents a challenge known as the cocktail party problem. Machine learning has tackled similar issues in human speech recognition. For instance, a model called Deep Karaoke was trained to separate vocals from instruments in music tracks and is now being applied to animal sounds, enabling researchers to isolate specific calls from groups of animals.

Researchers are also exploring AI's role in enhancing playback experiments. For example, models can be trained to generate unique versions of a sound after analyzing numerous recordings. This type of supervised learning, where models learn from human-labeled data, has limitations; it relies on existing knowledge of animal communication, which is still limited.

Yossi Yovel, a researcher who trained a model on 15,000 vocalizations from Egyptian fruit bats, found it could identify the caller, the context, and the recipient of the call. However, he acknowledged that human bias might overlook critical nuances in bat communication.

This limitation has led some AI researchers to advocate for self-supervised models, similar to those used in natural language processing. These models can sift through large amounts of unlabeled data, identifying patterns without human input. Aza Raskin, co-founder of the Earth Species Project, believes that language has a shape defined by relationships among words. This concept led to a groundbreaking discovery in 2017: the structure of different languages could be aligned, suggesting that similar words occupy the same conceptual space across languages.

Raskin and his team hope to apply this concept to animal communication, enabling translations without needing a “Rosetta Stone.” However, this task is complicated because animals communicate not only through sound but also through other senses. Yet, advancements in image generation models show potential for bridging the gap between sound and visual communication.

While researchers remain optimistic about interspecies communication, there are challenges, particularly in validating AI models. How can we assess their accuracy in understanding communications so distinct from our own? We also risk setting unrealistic expectations about conversing with nonhuman animals using shared language.

To build effective models, a comprehensive database of animal sounds is crucial. Researchers worldwide are undertaking extensive data collection efforts, tagging and recording animal vocalizations to create a rich resource for these models.

Ultimately, while it remains uncertain if true interspecies communication through AI will become a reality, researchers are hopeful that these discoveries will deepen our understanding and appreciation of the animals we share the planet with. They remind us that we are not alone in our capacity to communicate, care, and reflect on our existence. All species deserve recognition and respect in the tapestry of life on Earth.