If you’ve ever heard a bird sing and wondered how you could determine what type of bird it might be without laboriously searching through the records, the Cornell Lab of Ornithology has you covered.
The lab has recently improved its Merlin application for smartphone, designed for new and experienced bird watchers. It now has an AI-infused “Sound ID” feature that can capture bird sounds and compare them to crowdsourced samples to determine which bird is making that sound. Since the feature launched at the end of last month, it has become the most popular feature of the app (which also offers AI tools to identify birds in photos), and people have used it. to identify over a million birds. The number of new users has also increased by 58% from the two weeks before launch, and 44% from the same period last year, according to Drew Weber, project coordinator at Merlin.
“Users are finally putting a name on the birds that they have seen or heard before, but have never been able to identify them, and are really, very excited about it,” Weber wrote in an email, explaining that it is often easier to get a good audio sample from a bird. than a good photo.
Even when listening to bird sounds, the app still relies on recent advances in image recognition, says project research engineer Grant Van Horn. When you ask the app to record sounds around you and analyze them for bird calls, it actually turns the sound into a visual graph called a spectrogram, similar to what you might see in an audio editing program. Then he analyzes this spectrogram to look for similarities to known bird calls, which come from the Cornell Lab. EBird citizen science project.
The objective is to identify many audio samples without generating false identifications. But some birds are easier to identify than others, Van Horn says.
“Some just don’t have a lot of variation,” he says. “Other birds are much more difficult, either they have a much larger repertoire that can change dynamically, or they are actually imitators. “
Birds like blue jays and mockingbirds that mimic the sounds of other birds are naturally more difficult to identify conclusively, but the team has ways to improve the app around difficult birds. When there are problems identifying particular types of calls, they can search for additional samples of that bird, have an expert confirm that they are indeed properly classified, and add them to the data set. training.
The app isn’t the only one that works as a sort of Shazam for birds, but it’s completely free and assures users that it doesn’t submit their audio to a central server, although Cornell may offer the option. to share samples in the future. Instead, all processing is done on users’ iOS or Android devices, which both protects privacy and ensures that people can use the app while on hikes or in other places with cell reception. limited (although you must let the app download a dataset for your region the first time you use it in a particular region).
“Currently, no data is shared with Cornell,” says Van Horn. “Users don’t have to worry about privacy issues there. “
Right now, the team is working to further perfect the model before next spring, when bird watchers are likely to head to the parks and trails in hopes of identifying migratory birds en route north. . One challenge will be to ensure that the app can handle multiple overlapping bird calls at a time when birds will be particularly abundant. Cornell designers will also continue to work on managing birds that the app struggles to recognize. This includes a species – abundant near Van Horn’s original base in Ithaca, New York – that he believes could be so common in the context of other bird records that the AI has actually taught the ignore.
“It takes a long time for the app to make a suggestion on red-shouldered robins,” he says, “and that’s something I’ll keep repeating and trying to improve.”