Google is using AI to help humans talk to dolphins

These days, it seems like no worker is safe from an AI replacement—even Dr. Dolittle.

Google’s AI research lab DeepMind yesterday unveiled a new AI model aimed at helping scientists translate dolphin vocalizations, with an end goal of allowing for basic communication between humans and dolphins.

Background: Dolphins are some of the smartest and most communicative animals on Earth, with social interactions so complex that researchers at the Wild Dolphin Project (WDP) have spent the last 40+ years attempting to decipher them.

In the process, WDP has amassed decades’ worth of dolphin audio and video, using it to tie certain sounds to behavioral aspects like courtships, unique names, and dolphin squabbles.

That’s where AI comes in. Using WDP’s massive dataset, Google researchers created a generative AI model called DolphinGemma that uses a special audio technology to tokenize dolphin vocalizations.

• DolphinGemma operates on an audio-in, audio-out model. So after being provided with a dolphin vocalization, the model does what human-centric language models like ChatGPT do—it predicts the next token.
• Once the model is completed and fine-tuned, those predicted tokens would ideally be sounds that a dolphin can understand.

Looking ahead…Researchers hope DolphinGemma will help tease out complex patterns, and eventually allow them to create a shared vocabulary between dolphins and humans to enable basic interactions.