It’s early May, and I stand at one of my favorite birding corners half a mile from my house. American Robins chirrup like crazy, and the calls of Canada Geese, a Northern Flicker, and a Song Sparrow come at me from different directions. But more birds are out there. I can hear them even though my crappy ears and aging brain can’t parse what they might be. “Ah,” I think. “Let’s try the app.”
I pull out my phone, punch a couple of buttons, and record a short snippet of sound. Then I press “Analyze.” After a delay of several seconds, two species’ names appear. The first is Cedar Waxwing. That doesn’t surprise me since I can look up and see six of them perched in a tree. The second name does surprise me: Yellow-rumped Warbler. I haven’t been able to detect that one, but now that I’m alerted to its presence, I listen specifically for it and — shazam — its sweet spring song quickly penetrates my awareness. The discovery puts a bounce in my step, and not just because I’ve heard the warbler. It’s because I feel myself standing at the threshold of dramatically new birding possibilities.
The app I’ve just tested is called BirdNET, and its purpose is simple: to use the vocalizations of birds to identify them. Birdsong apps, of course, are not new. In my February 2021 BirdWatching article on song apps, I reviewed several designed to help birders recognize what we’re hearing around us. Two things set BirdNET apart from other birdsong apps, however. One is how it analyzes bird vocalizations. The second is the astonishing degree of accuracy that it achieves. In the parlance of modern business, these features establish BirdNET as a disruptive technology for birders and scientists alike.
BirdNET is the brainchild of Stefan Kahl, a postdoctoral student of media informatics at Germany’s Chemnitz University and researcher at the K. Lisa Yang Center for Conservation Bioacoustics at the Cornell Lab of Ornithology. “I was doing my master’s degree in computer vision,” he recalls, “and we had this project to identify acoustic events. So I was mostly thinking about using sounds in assisted living situations to see if we could detect people falling down or needing help with anything. Unfortunately, we had problems gathering enough data because, you know, what does it sound like when somebody is actually tripping? Sounds were really hard to authenticate. After all, you don’t want to detect a problem just because someone turns on a TV.”
Even though Kahl and his team were trying to identify acoustic events, their actual computer challenge was visual. That’s because to analyze a sound, a program must first convert it to a graphical picture called a spectrogram — basically, a chart of its frequency and pattern. Training the computer algorithm then requires feeding it hundreds, if not thousands, of known sounds, but obtaining enough of these known samples can pose a problem. Through a chance meeting on a research cruise, however, one of Kahl’s colleagues, Marc Ritter, met Holger Klinck, who would soon become the director of Cornell’s Yang Center. It was a meeting that would help launch Kahl’s work in an exciting new direction.
“I didn’t have a particular interest in birds at the time,” Kahl remembers, “but that quickly changed when Holger introduced me to the field of avian bioacoustics in 2016. Holger invited us over to Ithaca, and I noticed right away that with the high number of recordings that were online for birds, we were able to train better algorithms, and we could actually have an impact on people because such an algorithm was in high demand.”
Kahl and his team felt confident enough about their prototype that they entered it in BirdCLEF 2017, a competition to use computer programs to identify bird recordings. Their program scored high marks. “But I also knew that we could improve our systems even further,” Kahl continues, “and that’s when I decided to dedicate my PhD dissertation to the recognition of bird sounds.”
The BirdNET program is what’s called a deep artificial neural network, and it operates differently from standard computer programs. Traditionally, most computer programs for identifying bird vocalizations have “manually” measured note frequency, pattern, duration, and other variables and compared them to known parameters to make a prediction. Kahl explains that while these kinds of programs do work for some species, they quickly run into limitations. “There’s only so much you can do with a limited number of features,” he explains. “Birds are just too diverse.”
Instead, Kahl’s program uses neural networks, often referred to as AI or artificial intelligence. This kind of algorithm differs from older approaches because it teaches itself. “In principle, the program is still comparing a spectrogram with known samples because you need a lot of playing samples to learn to identify a complex pattern,” Kahl says. “But these patterns are not just note sequences or element sequences, and the program doesn’t match a frequency range or something. It’s more than that. Really, we don’t exactly know what the program is doing. It’s very complex.
“Sometimes people say it’s a black box,” he continues, “and in some sense, that’s true because you really don’t know what these kinds of programs are learning and why they are learning it. There is certainly some part of the bird call that BirdNET focuses on more than others, but it’s like when you as a human recognize a new song from your favorite band that is playing on the radio for the first time. You instantly know the band even though you’ve never heard the particular lyrics, notes, and melody of the new song. You recognize the band because you’re familiar with what it sounds like, and I think that’s the best way of thinking about how this algorithm works. It just learns these features extracted from hundreds of thousands of recordings.”
To teach the program, Kahl and his colleagues feed it recordings with species labels from Cornell’s Macaulay Library, the bird-sounds website Xeno-canto, and other sources. Just how many samples the program needs to learn to identify a raw, unlabeled bird vocalization depends on the complexity of the song. “A chickadee is very simple,” Kahl says, “so I would say the program needs a few hundred examples or a thousand to learn that one. For most species, it needs between 500 and 3,000 samples, but there is an upper limit where the program doesn’t improve even if you throw more samples at it.”
Making the program effective, though, requires teaching it more than bird calls. Because the real world is full of many other noises, BirdNET has to learn a whole symphony of sounds of people, animals, traffic, running water, and more.
One hilarious problem occurred when trying to teach BirdNET to identify owls. “If you listen to most recordings people have made of owls,” Kahl explains, “you hear the owl, and then you hear people say, ‘Oh, there’s the owl! There it is!’” As a result, BirdNET learned that an owl wasn’t an owl unless human voices accompanied it. A similar thing happened with the sounds of human footsteps for owls and other species. This forced Kahl’s team to teach BirdNET to distinguish between bird and multiple human noises.
Despite the challenges, Kahl’s team quickly taught BirdNET more than 1,000 species from Europe, the U.S., and Canada — places with rich libraries of recordings for the program to learn from. Even more astonishing, the program learned to accurately identify these species with about 80 percent accuracy, eclipsing any other bird vocalization tool available.
Once they had the program working, the team set out to make BirdNET available from both browsers and mobile devices. The first BirdNET app was released in October 2018 for Android devices because of its open architecture. iOS users finally got their version in December 2020, helping to usher in a revolution for both birders and scientists.
“If you listen to most recordings people have made of owls, you hear the owl, and then you hear people say, ‘Oh, there’s
the owl! There it is!’
Leveling the Birding Playing Field
One thing that distinguishes exceptional from less-vaunted birders is our ability to hear and identify bird calls. The deeper you get into birding, the more you realize how sneaky birds are and that the only way to detect many of them is through their vocalizations — something that is especially true of forest and grassland species.
The problem for many of us is that we don’t even begin dedicated birding until long after our physical and mental abilities have peaked. Auditory function is a common casualty of living a long life, and by the time we reach our 50s or 60s, many of us have suffered significant hearing loss. Hearing aids help, but they aren’t a panacea as they can distort birdsong and directionality while still failing to pick up many high frequencies.
Even learning new birdsong poses another challenge. While the typical 15-year-old might absorb new species calls as fast as a flicker sucking up suet, we older birders often must hear a bird dozens or hundreds of times to learn it — and may then have to learn it anew each year! For birders in either of these two categories, BirdNET allows us to greatly expand our birding horizons.
In my first few weeks using BirdNET, I quickly discovered that it helped me both detect species I didn’t realize were nearby and allowed me to quiz myself on what I was hearing. This second aspect has been especially fun to explore. In spring, for instance, I often get befuddled by the barrage of warbler-type songs around me. With BirdNET, I can hear something and guess, “Ah, that’s a Ruby-crowned Kinglet,” or “I think that’s a Nashville Warbler.” Then, I record a snippet and have the app analyze it. It doesn’t always hit the target — especially in conditions roaring with background noise — but more times than not, it lets me know if I’ve made a mistake.
Opening Up the Scientific Process
As exciting as it is for individual birders, BirdNET as a scientific tool promises to demolish stubborn barriers to ornithological research. Connor Wood is a postdoctoral fellow in Cornell’s Yang Center. “I think one of the things that’s most exciting to me about the BirdNET app,” he shares, “is that it opens up the scientific process to vastly more people than almost any other citizen science program that I’m aware of.” By the summer of 2020, BirdNET already had more than a million active users in North America and Europe — and that was months before the iOS version got released. What’s more, the number barely dropped through the following winter. “What that means,” Wood says, “is that we really have the potential to have a global network of biodiversity sensors.”
One of Wood and Kahl’s first scientific priorities has been to test BirdNET against other ornithological studies. A 2020 study released by Canadian scientists, for example, examined the spread of a new song variety of the White-throated Sparrow. Traditionally, the bird’s song terminated in what the scientists call a “triplet ending,” but since the year 2000, a “doublet ending” had rapidly spread from west to east across North America — to the point that, according to the Canadian study, the doublet had almost completely dominated populations in the West and Midwest.
The study was based on hundreds of recordings that the scientists collected themselves between 2000 and 2019, and on additional data from other sources, but Wood and Kahl wanted to see if they could duplicate the result using only calls submitted by BirdNET users in 2020. Digging into the BirdNET database, they selected high-quality White-throated Sparrow songs from across the continent to see if they had triplet or doublet endings. Like the Canadian researchers, they found both varieties in BirdNET submissions, but the BirdNET data did not show the doublet totally displacing the triplet variety. Instead, BirdNET recordings revealed that both varieties continue to exist in most regions of the continent (see map) — a surprising and eye-opening result.
In another study, Wood and Kahl used BirdNET submissions to confirm migration routes of Common Cranes from Spain to northern Europe and back. BirdNET data generated a beautiful confirmation of the birds’ main route, and it also lent support for alternative southern routes that were only first described in 2013 by a team of Italian scientists. These routes are particularly exciting because increasing numbers of cranes appear to be using them, possibly reflecting the overall rebounding crane population and new behaviors associated with climate changes.
Just as important, both the sparrow and crane examples give a tantalizing glimpse of the many ways BirdNET may soon broaden our understanding of bird biology and behavior — and give scientists, governments, and conservation workers essential new information to protect birds.
Listening Around the World
Though BirdNET can currently identify only about 10 percent of earth’s species, Kahl and his colleagues are working to raise the total number to more than 5,000, giving the app a reach far beyond Europe and North America. This will not only help birders on every continent, but also it will expand the scientific applications of the app. “Once we get this increasingly global buy-in of app users,” Wood excitedly explains, “we can start to find patterns and places outside of Europe and North America, where there’s not been heavy investment in western scientific research and citizen science programs. I think that’s when app users can really contribute novel science. In the case of flyways, we can potentially do some cool targeted conservation intervention, identifying hotspots and migratory routes.”
Meanwhile, Kahl and his team are working to improve BirdNET in other ways. One goal is to make it a stand-alone program. Currently, the app must send recordings back to mainframe computers for analysis, making it necessary for birders to have an internet connection or save recordings for later submission. Kahl’s team, though, has a working stand-alone prototype that can analyze a three-second recording in about two-tenths of a second — something it hopes to release in the next year or so. It also is developing a continuously running version of the app that can be used to conduct point counts of the type that scientists frequently use to monitor bird populations.
As word of BirdNET spreads, Kahl is receiving myriad requests from birders and scientists to add other capabilities — including a Pokémon-Go-type feature that allows users to “collect” bird songs of various species during their birding outings. All of this will take time.
“There are a lot of technical challenges that we have to figure out,” Kahl explains. “We’re just a small team. There is no big company behind it. It’s basically me and a few researchers helping me.”
The small size of the team, in fact, makes it even more remarkable what it has accomplished. Even without rolling out new features, Kahl and his colleagues have ensured that BirdNET will be an important part of the ornithological landscape for years to come. Which leaves only one question: Have you downloaded your copy yet?
BirdNET data generated a beautiful confirmation of the birds’ main route, and it also lent support for alternative southern routes that were only first described in 2013 by a team of Italian scientists.
BirdNET is available for free through the Google Play and Apple App stores. To learn more about it and explore some of its data, visit birdnet.cornell.edu. To support BirdNET developers and ongoing research, visit birdnet.cornell.edu/donate.
Sound ID for Merlin
In a case of convergent app evolution, the Cornell Lab of Ornithology recently unveiled a new song identification feature for its excellent Merlin Bird ID app. Like BirdNET, Merlin’s “Sound ID” feature was developed with AI technology, using known spectrograms to teach it to identify more than 400 North American species. Once you start Sound ID, a continuous spectrogram scrolls across the screen, and when a bird is identified, the species name pops up below, accompanied by a photo.
Especially for someone with a hearing disability or who just doesn’t know birdsong, it can be like magic watching a succession of bird names appear before your eyes. No doubt it will make a terrific teaching tool, especially for those just getting into birding.
The program is not without its limitations. Unlike BirdNET, Merlin’s Sound ID is a stand-alone feature, which gives it the flexibility to work without cell service, but it may also limit its accuracy. In quick tests around my largely forested neighborhood, Sound ID correctly picked out most species but mistook some common birds for unlikely rarities, such as American Three-toed Woodpecker and Northern Pygmy-Owl. Still, the app constitutes a major step forward in birding technology, and Cornell will undoubtedly continue to improve it over time. For now, though, birders will probably want to use Sound ID more as a learning tool and continue to hone and rely on other skills to confirm their own bird observations.
This article first appeared in our September/October 2021 issue.