Photo credit: www.technologyreview.com

Advancements in Acoustic Ecology: BirdVox Project Uses Machine Learning to Decode Nocturnal Flight Calls

In the late 19th century, researchers made a groundbreaking discovery: migratory birds produce distinct flight calls at night, akin to “acoustic fingerprints.” This revelation laid the groundwork for further research into avian communication. The commercialization of microphones in the 1950s allowed scientists to begin recording birds after sunset, leading to invaluable acoustic ecology studies. In the 1990s, leading researcher Farnsworth actively contributed to this field. However, identifying the intricate and often subtle calls remained a daunting task, as many of them exist within the upper limits of human hearing. As a result, scientists accumulated vast amounts of audio recordings that had to be analyzed in real-time through spectrograms, tools that visualize sound. Despite advances in digital recording methods making the collection of data easier, Farnsworth highlighted a persistent challenge: the increasing difficulty in analyzing the growing volumes of audio data collected.

A pivotal moment came when Farnsworth collaborated with Juan Pablo Bello, the director of NYU’s Music and Audio Research Lab. Bello, who had recently applied machine learning techniques to identify sources of urban noise in New York City, was intrigued by the potential to tackle the challenge of nocturnal bird calls. He assembled a team, including French machine-listening expert Vincent Lostanlen, which led to the establishment of the BirdVox project in 2015. “Everyone believed that once this issue was resolved, it could unlock a wealth of information,” Farnsworth noted. Nonetheless, Lostanlen reflected on the initial skepticism: “At first, there were no indicators that this was achievable.” The prospect of using machine learning to replicate the auditory discernment of seasoned ornithologists like Farnsworth seemed daunting.

Bello expressed profound admiration for Farnsworth’s expertise, stating, “Andrew is our hero. Our goal is to mimic his capabilities with technology.”

To initiate the project, the team focused on training BirdVoxDetect, a neural network designed to filter out noise interference, such as the sounds produced by rain-damaged microphones. The next challenge was to train the system to recognize the specific flight calls of various bird species. This task was complicated by the fact that these calls can be easily mistaken for other noises, similar to how a smart speaker might struggle to identify its wake word amid background sounds. Unlike smart devices that utilize distinct trigger phrases, the researchers faced an additional complexity: the diverse range of bird calls shaped by evolution. “Charles Darwin effectively made that choice for us,” Lostanlen quipped. Fortunately, the team possessed an extensive corpus of training data, thanks to Farnsworth’s group, which had meticulously hand-annotated thousands of hours of recorded sounds from microphones situated in Ithaca.

Once BirdVoxDetect was capable of recognizing flight calls, the project encountered another significant hurdle: classifying these calls by species—a task even expert bird watchers find challenging. To address this uncertainty and the absence of training data for each species, the researchers implemented a hierarchical classification system. For instance, when analyzing a call, BirdVoxDetect may successfully determine the bird’s order and family, even if the exact species remains unclear, mirroring the way a bird watcher might identify a call as belonging to the warbler family, regardless of whether it is a yellow-rumped or chestnut-sided warbler. During training sessions, the neural network received less penalty for incorrectly grouping species that were closely related in the taxonomic hierarchy.

Source
www.technologyreview.com