Global AI models struggle with Indian wildlife sounds. 59 volunteers built a fix

Last year, a group of ecologists and enthusiasts embarked on a mission to capture the sounds of India's diverse ecosystems. They collected nearly 100 hours of recordings spanning various terrains, from the Western Ghats to the East Deccan's deciduous forests. The recordings featured a mix of bird calls, insect chirps, frog croaks, bat and reptile noises, and even marine sounds.

After meticulous analysis, researchers cataloged the species behind each sound, marking their presence and frequency on spectrograms. On July 21, the dataset was published on bioRxiv, an online server where scientists share their pre-publication work. This marks India's first open-access, crowdsourced ecoacoustic dataset of multiple animal groups, a treasure trove of wild sounds built by volunteers and freely available for use.

With 518 species represented across 25 states and union territories, the dataset is a significant step towards using sound to study nature in India. Previously, ecoacoustic recordings were scattered across various collections. This open-access dataset aids scientists in monitoring biodiversity and enhances AI's ability to identify Indian wildlife accurately.

Dr. Pooja Choksi, an ecological restoration scientist and co-founder of the Indian Ecoacoustics Network (IEN), explains that recordings offer a comprehensive snapshot of an ecosystem's soundscape, capturing species that might go unnoticed through visual observations alone. AI models like BirdNET and Perch can automatically identify species from these recordings, simplifying the analysis process.

However, a gap exists in India. Machine learning models for acoustic species recognition are largely trained on data from the Global North. This lack of representative data from Indian ecosystems leads to poor performance of these models in the country's tropical regions. The IEN aims to address this issue by providing a centralized, labeled dataset of high-quality acoustic data from diverse Indian ecosystems.

The dataset, which credits 59 contributors, acknowledges the need for noisy, representative data to improve AI accuracy in identifying species amidst background noise and environmental disturbances.