In a remarkable stride towards advancing wildlife monitoring, a team of researchers at the University of Moncton in Canada has unleashed ECOGEN, an unprecedented deep learning AI marvel. This groundbreaking tool specializes in crafting authentic birdsongs, not for the birds themselves, but to revolutionize the training of bird identification tools used by ecologists. Unveiled in the British Ecological Society journal, Methods in Ecology and Evolution, the study introduces ECOGEN as a solution to the conundrum faced by conservationists and ecologists when dealing with the identification of rare avian species.
Unlike routine bird identification aided by numerous apps and software, recognizing rare birds poses a unique challenge. ECOGEN steps into this void by generating lifelike bird sounds, particularly catering to species with limited or scarce audio recordings. This tool, a trailblazer in its category, enriches the training dataset for audio identification tools crucial for ecological monitoring.
According to Dr. Nicolas Lecomte, one of the lead researchers, “Due to significant global changes in animal populations, there is an urgent need for automated tools, such as acoustic monitoring, to track shifts in biodiversity.” ECOGEN plays a pivotal role by creating new instances of bird sounds to bolster AI models, especially for rare, elusive, or sensitive species. This innovative approach expands the sound library without further disturbing the animals or necessitating additional fieldwork.
The impact of ECOGEN is evident in its application; when integrated into a birdsong identifier, it elevated bird song classification accuracy by an average of 12%. The tool’s potential extends beyond birds, offering promise for mammals, fish, insects, and amphibians—a testament to its adaptability.
Notably, ECOGEN stands out not just for its efficacy but for its accessibility. Its open-source nature and compatibility with basic computers make it a democratizing force in the realm of conservation technology.
The process involves the transformation of real bird song recordings into spectrograms, providing visual representations of sounds. ECOGEN then crafts new AI images based on these spectrograms, effectively amplifying the dataset for rare species with scant recordings. The resulting spectrograms are translated back into audio, serving as a novel and dynamic approach to train bird sound identifiers. The study harnessed a robust dataset of 23,784 wild bird recordings, spanning 264 species from diverse corners of the globe.