Understanding bird calls using AI and ML

The Coller-Dolittle Challenge for Interspecies Two-Way Communication awarded University of California, Berkeley scientist Dr. Julie Ely an annual prize of $100,000. With the help of machine learning and AI, he has built a dictionary of 11 words used by zebra finches.

Inspired by the Turing Test, this award recognizes the most promising research aimed at developing algorithms for communicating with animals. The challenge is being carried out under the auspices of the Jeremy Koller Foundation, a UK-based charity affiliated with Israel’s Tel Aviv University, and its name is a combination of the foundation’s founder, Jeremy Koller, and Dr. Dolittle, a fictional character who can talk to animals.

Launched in 2024, the initiative features an annual progressive prize of $100,000 and a grand prize of $10 million in equity or $500,000 in cash to the team that successfully creates truly independent, two-way communication in which animals interact with humans without realizing they are interacting with humans.

To win the annual award, entries must meet rigorous scientific criteria. Research must non-invasively decipher or interface animal communication, demonstrate understanding or interaction across multiple situations (e.g. foraging, mating, alarm calls) using the organisms’ own natural signals, and demonstrate measurable responses from organisms when these signals are broadcast to them.

In 2025, the inaugural award was awarded to a research team led by Dr. Laela Sayegh of Woods Hole Oceanographic Institution who used AI to analyze and decipher the distinctive whistle types and communication patterns of wild bottlenose dolphins.

This year, the award went to Dr. Julie Yee of the University of California, Berkeley. That’s because her 15 years of research created a core “lexicon” of 11 words for zebra finch calls and demonstrated that birds group calls based on behavioral meaning rather than simply acoustic similarity.

Zebra finches sing continuously, producing thousands of hours of dense audio data. Traditional manual audio analysis would have taken years to classify this volume, but AI algorithms parsed these vast datasets at high speed, systematically detecting, separating, and tracking individual acoustic events across different contexts. MacA Khinka learning model was employed to extract and classify the precise acoustic features of zebra finch calls, and to map the structural differences between the 11 types of calls that make up the zebra finch ‘lexicon’, an AI tool revealed complex bioacoustic patterns containing subtle variations that are difficult to distinguish with human hearing.

Artificial intelligence and machine learning were the basis for creating the dictionary, but it was classic behavior that was used to test the results and confirm the meaning of sounds obtained in experiments where birds were trained to interact with a playback system. By mapping the precise acoustic distances calculated by the AI against the birds’ behavioral responses, the researchers noticed a significant anomaly. This means that the bird classification error does not match the acoustic similarity. Although the birds regularly confused acoustically distinct calls that shared similar behavioral meanings (such as long-range and short-range contact calls), they readily distinguished very similar sounds that meant completely different things, leading to the conclusion that birds group calls based on behavioral meaning rather than mere acoustic similarity.

In this TV news interview, Julie Ely describes an experiment she devised and goes on to explain how being able to talk to birds is beneficial to them.

Detailed information

Koller Foundation Announcement

Researchers use AI to decipher dog language

Project SPARROW starts

Cat to human translation app

To stay informed about new articles in I Programmer, sign up for our weekly newsletter, subscribe to our RSS feed, and follow us on Facebook or Linkedin.