Google's New AI Model DolphinGemma Can Communicate with Dolphins

In "The Hitchhiker’s Guide to the Galaxy," dolphins tried to warn humans that Earth was about to be destroyed. They performed all sorts of acrobatics, jumping and flipping, trying to convey the message, "So long, and thanks for all the fish!" But humans misunderstood this crucial message, thinking the dolphins were just being playful. Eventually, as Earth faced demolition to make way for a "hyperspace bypass," the dolphins gave up on communication and left, leaving humans to face the impending doom alone.

Today, this tragedy won't happen, thanks to Google's research team and DeepMind scientists who have introduced DolphinGemma, a language model specifically designed to process dolphin vocal sequences. Finally, we won't misinterpret dolphin warnings (if they exist), and if the world is ending, remember to follow the dolphins! (Just kidding!) Now, we might even be able to talk to animals like dolphins and understand their messages.

How Did Scientists "Decode" Dolphin Language?

Dolphins are incredibly intelligent and expressive, which has long intrigued scientific teams aiming to "decode" their language. One such group, the Wild Dolphin Project (WDP), has been observing wild Atlantic spotted dolphins in the Bahamas for years. Founded by Dr. Denise Herzing, the team collaborated with Georgia Institute of Technology in 2012 to develop the CHAT (Cetacean Hearing Augmentation Telemetry) system.

CHAT is an underwater interactive system designed for "simplified two-way communication" with dolphins. The goal isn't to understand dolphin language but to create a set of synthesized human sounds (whistles) paired with objects dolphins frequently encounter, like seaweed, seagrass, or research equipment. The hope is that dolphins, out of curiosity, will mimic these sounds. Once they successfully convey the intent of "give me this," researchers immediately provide the item, reinforcing the learning and establishing a shared vocabulary. This system is a step towards being able to talk to animals, starting with dolphins.

The WDP team has meticulously tracked underwater audio and video of many dolphins, documenting each dolphin's identity, behavior, social interactions, and life history. They can even correlate specific sounds with behavioral contexts, like a mother dolphin's "signature whistle" to call her calf. These rare and valuable dolphin datasets form the foundation of DolphinGemma.

The Wild Dolphin Project, founded in 1985, has been trying to communicate with dolphins for 40 years. (Source: WDP)

What is DolphinGemma?

Enter the era of AI. Gemma is Google's open-source lightweight LLM model, designed to run on edge devices like smartphones. As with the Transformer architecture, AI doesn't truly understand language but can analyze patterns to make predictions. If it can "decode" human language, it can naturally "decode" dolphin language too. 

So, Google used these long-collected dolphin datasets to train this dolphin language model, employing its SoundStream audio processing technology to create DolphinGemma, an AI model that takes audio input and produces audio output. In short, it can receive dolphin sounds, analyze their structure and patterns, predict subsequent sounds, and simulate conversations with dolphins, gathering more data in the process. What's even more exciting is that this model, with around 400 million parameters, is compact enough to run directly on a Pixel phone. This means that scientists can simply carry a customized Pixel 9 phone, equipped with a microphone and speaker, to have seamless underwater "chats" with dolphins, making it possible to talk to animals in a way never before imagined.

Indeed, DolphinGemma currently only has data for Atlantic spotted dolphins, and other dolphins might want their language models too. Therefore, Google plans to open-source DolphinGemma, allowing other scientists to use this model to analyze dolphins from different regions and fine-tune it, benefiting dolphins worldwide. Perhaps in the near future, not only will we have LLMs that truly decode dolphin language, but we might also have language models for cats and dogs, making pet communication a reality, and enabling us to talk to animals across various species!