Understanding the language of dolphins has always been a very difficult and fascinating challenge for scientists between whistles, impulses, very sharp verses and, sometimes, rhythmic. But what would happen if we could not only listen to the dolphins, but also decode their communicative patterns to the point of generating realistic answers? On the occasion of the National Dolphin Day, Google announces together with researchers of the Georgia Institute of Technology and al Wild Dolphin Project (WDP) important progress on DolphinGemma: a new generation artificial intelligence model, designed to learn the structure of dolphins vocalizations and generate new sound sequences similar to real ones. A real turning point in the research on interspecies communication, which leads the Ai to curl still unexplored territories of our link with the marine world.
A project based on decades of observations
It is a project, the one dedicated to dolphins, which starts from afar. And a long post published on the blog of the American tech giant explains it. Really understanding a species requires deep context, and this is exactly what WDP offers. Since 1985, this project has been carrying out The longest -running underwater research on dolphins in the worldby studying a community of macular dolphins of the Atlantic (Stella Frontalis), around the Bahamas, through multiple generations. Their approach is not invasive with the aim of moving ‘step by step’ by entering the tip of their world and “with their conditions”. The result, however, is an incredibly rich and colorful dataset: decades of video and diving audio recordings, combined with the identity of individual dolphins, their life stories and the observed behaviors. Stories, many, in -depth.
The help of Dolphingemma
Analyze the natural communication of the dolphins – complex, dynamic and full of nuances – is as said a titanic challenge. To do this, scientists must refine their techniques, update their tools, to experience continuously. Thanks to the labeled dataset of the WDP, Google was able to train Dolphingemma, an AI model that uses advanced audio technologies and that accelerated the study process. Specifically, use the so -called tokenizer SoundStream To efficiently represent the sounds of the dolphins, then tried through an architecture designed to manage complex sequences. The model, with about 400 million parameters, was optimized to work directly on smartphone Pixel used by the WDP during diving.
Dolphingemma was born in the wake of the Gemma family, the collection of Google’s open-source models, based on the same research that feeds the well-known Gemini models. It is an audio-in model, audio-out: it analyzes the natural sequences of dolphin sounds, identifies patterns and structure, and predicts the most likely sound that it will follow, just as linguistic models do with human words.
A new shared language?
During this research season, Dolphingemma will be used directly on the field. His contribution will be immediate: Identify recurrent patterns, sound clusters and reliable sequenceshelping researchers to discover structures and possible meanings in the natural communication of dolphins. A task that, as Google recalls, has so far required years of human work. But there is more: researchers are also supporting Dolphingemma with synthetic sounds created specifically to refer to Objects loved by dolphins (such as algae or gaming accessories). The long -term goal is to develop a shared vocabulary to facilitate a real communicative interaction between man and dolphin.
In addition to the analysis of natural communication, WDP also has another task: to develop bidirectional interaction in real time through technology. Thus was born the system CHAT (Cetacean Hearing Augmentation Telemetry), made with the support of the University of Georgia Tech. Chat is a diving computer designed not to directly decipher the natural language of the dolphins, but to build a shared and simple vocabulary, based on artificial whistles. The system associates each object (such as sargassi algae, marine herbs or colored scarves used by researchers) a unique synthetic whistle, different from the natural sounds of the dolphins.
By showing how these sounds are connected to interesting objects, researchers hope that the dolphins, always extremely curious, learn to imitate those whistles to make certain requests and embark on a form, even sketched, of communication. As the understanding of natural sounds grows, these can also be integrated into the chat system, making the line between listening and conversation more and more nuanced.
Click here and sign up for our WhatsApp channel! The news, in real time, of the Italy Agency now also on your smartphone