Read our COVID-19 research and news.

Jan Meeus/Unsplash

Watch out, birders: Artificial intelligence has learned to spot birds from their songs

Bird populations are plummeting, thanks to logging, agriculture, and climate change. Scientists keep track of species by recording their calls, but even the best computer programs can’t reliably distinguish bird calls from other sounds. Now, thanks to a bit of crowdsourcing and a lot of artificial intelligence (AI), researchers say they have something to crow about.

AI algorithms can be as finicky as finches, often requiring manual calibration and retraining for each new location or species. So an interdisciplinary group of researchers launched the Bird Audio Detection challenge, which released hours of audio from environmental monitoring stations around Chernobyl, Ukraine, which they happened to have access to, as well as crowdsourced recordings, some of which came from an app called Warblr.

Humans labeled each 10-second clip as containing a bird call or not. Using so-called machine learning, in which computers learn from data, 30 teams trained their AIs on a set of the recordings for which labels were provided and then tested them on recordings for which they were not. Most relied on neural networks, a type of AI inspired by the brain that connects many small computing elements akin to neurons.

At the end of the monthlong contest, the best algorithm scored 89 out of 100 on a statistical measure of performance called AUC. A higher number, in this case, indicates the algorithm managed to avoid labeling nonbird sounds as bird sounds (humans, insects, or rain often threw them off) and avoid missing real bird sounds (usually because of faint recordings), the organizers report in a paper uploaded to the preprint server arXiv. The best previous algorithm they tested had an AUC score of 79.

The algorithm atop the pecking order could even generalize well enough to score 84 on samples of nocturnal bird calls that were very brief and hard to analyze and very different from the training sounds. The algorithms can’t outperform humans (who were used to label the data in the first place), but machines can operate all day and night and don’t mind the rain. It’s only a matter of time before an AI hatched from this competition takes flight in the real world.