Considered some of the planet’s most intelligent beings, dolphins possess remarkable cognitive abilities, including cooperation, teaching, and self-recognition. Over the years, researchers have endeavored to decode the intricate series of clicks and whistles that dolphins utilize for communication. Progress in this area may soon be enhanced through a partnership involving Google’s open AI model and various Pixel smartphones.
Google has been integrating generative AI into its wide array of services, and its collaboration with the Wild Dolphin Project (WDP) exemplifies this trend. Established in 1985, the WDP has been examining a specific group of Atlantic spotted dolphins employing non-invasive methods to monitor their behavior. The team compiles video and audio recordings, complemented by detailed observations of the dolphins’ actions.
A primary focus of the WDP is to analyze dolphin vocalizations and their influence on social interactions. With decades worth of underwater recordings, the researchers have started to link certain behaviors with distinct sounds. For instance, signature whistles among Atlantic spotted dolphins function similarly to names, helping individuals locate one another, while “squawk” patterns are frequently observed during confrontations.
According to WDP researcher Denise Herzing, grasping the composition and trends in dolphin communication is key to understanding whether it can be classified as a language. “We do not know if animals have words,” she stated.
An overview of DolphinGemma
The ultimate aspiration is to establish a form of communication with dolphins, assuming such a linguistic structure exists. This ambition has driven the WDP to compile a vast, carefully categorized dataset, which Google identifies as ideal for analysis via generative AI.
Introducing DolphinGemma
The large language models (LLMs) that have become prevalent in consumer technology function by predicting patterns. Users input data, and the models generate subsequent tokens repetitively until an output is formed. When these models are effectively trained, the produced output can closely resemble human-created text. Google and the WDP now aim to adapt this methodology to create DolphinGemma, a model tailored for marine communication.
DolphinGemma leverages Google’s Gemma open AI models, which share the foundational technology of the company’s commercial Gemini models. This innovative communication model utilizes SoundStream, an audio technology developed by Google, to tokenize dolphin vocalizations, enabling sounds to be entered into the model in real-time as they are recorded.
