AI uncovers how birds remix their songs over time

If you thought only humans had unique musical tastes that differed drastically across geographies and generations, think again.

Researchers have long suspected that migration and population dynamics shape the musical repertoire of songbirds. Now, they have empirical evidence to back up this hypothesis.

A new study published in the journal Current Biology provides insights into how age, population dynamics and migration are key factors that determine how birdsong evolves over time. With the help of artificial intelligence models, the team behind the research found that “demographic variation affects vocal cultures” for great tits (Parus major) in Oxfordshire in the U.K.

“The process of song acquisition relies on who they’re exposed to, the movement of individuals and the age of those individuals within populations,” study lead author Nilo Merino Recalde, a postdoctoral biology researcher at the University of Oxford, told Mongabay in a video interview. “This leads to all sorts of really cool dynamics that are quite similar to what happens with human languages and music.”

Listening in on birdsongs and figuring out how they’re evolving could potentially help researchers understand how bird populations are faring. For instance, scientists might be able to correlate song patterns to a reduction or fragmentation of populations. The song types could also help them get a sense of how much the birds are moving around. “This is still speculative in a practical sense but all of the different processes that we care about when it comes to conservation are reflected in the songs,” Recalde said. “In principle, we should be able to infer things about how populations are doing based on sounds and songs.”

As part of the research, the team deployed acoustic recorders close to the birds’ nests in a woodland ecosystem. Since the bird population in the location has been studied and monitored for more than 77 years, the team already had access to a huge trove of data to get started with. “We knew about the comings and goings of different birds, who was born where and the ages of the birds,” Recalde said. “We built a layer of song recordings on top of all that data.”

Over the course of three years, the researchers collected 21,000 hours of audio data. Then came the task of identifying songs from the recordings and picking out the different notes that make up the songs. At the end of the process, the team ended up with more than 1 million notes that came from about 100,000 different birdsongs.

Once the data were classified, the team narrowed down the two questions they needed answers to: How similar were any two songs in the data set? And could they identify the birds based on the songs?

This is where they brought in deep metric learning, a discipline of AI that learns to measure similarities between data points. The team trained the model to be able to differentiate between different song types and repertoires. Using this, they were able to understand how many different song types there are in the area and how fast they change based on the population dynamics. They were also able to identify birds based on their songs with a high degree of accuracy. “Our main goal was to see if their movements and who they are around when they are learning really changes things,” Recalde said. “The main takeaway was that we were able to detect these, which is what the theory has predicted for a long time.”

Specifically, the team found that when a bird moves around a lot, the songs it ends up singing are more likely to be popular songs that are well-represented in the population. “What this means is that movement kind of homogenizes the song culture and the pool of songs,” Recalde said.

Conversely, birds that don’t move around as much create tiny pockets with unique songs that aren’t found anywhere else in the population. Birds that move into the area and are new to the population, the team found, don’t introduce songs from elsewhere. Instead, they expand their repertoire by learning songs that are already prevalent in the population. “Basically, what happens is that their songs are not, on average, different to those of the local birds, but tend to end up with larger repertoires,” Recalde said.

Aging was found to slow down the learning process. Older birds were found to have stopped learning news songs. Much like in humans, these birds act as repositories of old songs. “They know the songs of yore, some of which are no longer in style,” Recalde said. Mixed-age neighborhoods in the area had more song diversity with younger birds leading the adoption of new songs.

In short, the research confirmed age-old theories that the diversity of birdsongs evolves as a function of several factors. Recalde and his team are now working to understand how song changes are driven at a community level across space and time. The data they’ve collected so far are also available publicly for scientists anywhere to use and analyze for their own research.

“It’s very exciting to see that this data can open the doors to a lot more research along different lines,” Recalde said.

Banner image: While older birds were found to have stopped learning new songs, birds that moved around a lot tended to know the popular tunes. Image by David López Idiáquez.

Abhishyant Kidangoor is a staff writer at Mongabay. Find him on 𝕏 @AbhishyantPK.

Bioacoustics and AI help scientists listen in on elusive Australian cockatoos

Citation:

Merino Recalde, N., Estandía, A., Keen, S. C., Cole, E. F., & Sheldon, B. C. (2025). The demographic drivers of cultural evolution in bird song. Current Biology, 35(7), 1631-1640. doi:10.1016/j.cub.2025.02.016

Credits

Abhishyantkidangoor Editor

This story first appeared on Mongabay

This article is licensed under a Creative Commons Attribution-NoDerivatives 4.0 International License.

You may republish this article, so long as you credit the authors and Mongabay, and do not change the text. Please include a link back to the original article.