Scientists are teaching AI to listen to India’s forests

On July 21, the result was published on bioRxiv, an online server where scientists post studies while a journal reviews them — in this case Scientific Data, published by Nature Research. With it, India has its first open-access, crowdsourced ecoacoustic dataset of multiple distinct groups of organisms, a library of wild sound built by volunteers and free for anyone to use — a chorus of 518 species across 25 states and union territories.

Manipur Fulvetta

Set against the staggering scale of the subcontinent’s wildlife, the tally may be modest, but it is a step towards using sound to study nature in the country. Until now, India’s ecoacoustic recordings existed in scattered collections. This open-access dataset gives scientists a shared resource to improve biodiversity monitoring and helps AI identify Indian wildlife more accurately.

“When ecologists do bird surveys on the field, we usually do a point count,” said Pooja Choksi, a Mumbai-based ecological restoration scientist and a co-founder of the Indian Ecoacoustics Network (IEN), the collective behind the dataset.

“You go to a spot in a forest, wait for 15 minutes, and note down all the birds you can observe in the distance. But if you rely only on sight, especially in a dense forest, you’ll end up missing out on most of the wildlife,” she said.

“Recordings can be a perfect time capsule for how an ecosystem sounds like in time and space. They can capture more species, over frequencies the human ear can’t hear, note how they vocalise, and do it periodically over long periods of time. This makes acoustics a game changer for biodiversity monitoring in any ecosystem,” Choksi said.

Chiranjib Bora during a focal recording in Arunachal Pradesh. (Photo: Jibon Dutta)

Ecologists have used ecoacoustics to chart changes in a habitat over time, through conservation efforts or after tree felling. Tigers have been found to have vocal fingerprints, their chuffing unique to each. With AI, the arduous task of manually listening back has been simplified — deep learning models like BirdNET and Perch automatically detect the species heard. All of this is done non-invasively, with a nondescript recorder wrapped around a tree.

In all this, there is a gap for India. Kadambari Devarajan, a Mumbai-based independent researcher who uses data science for conservation, explained why. “Machine learning models for recognising species/ birds through acoustics are largely trained on data from the Global North,” she said.

“And as every ecosystem has its unique soundscape, made up of a mix of endemic and common species, background noises and weather interruptions, ambient noise, the models perform poorly in tropical regions like India,” she said.

The gap is not for want of effort. State governments, researchers and institutions have been using ecoacoustics, said Sarika Khanwilkar, a Pune-based conservation scientist, but their recordings sit scattered, most not labelled in the detail AI needs. An open, labelled dataset puts hours of ready-to-use audio in the hands of anyone who needs it. A field researcher mapping an ecosystem with BirdNET can feed this data into the model, instead of depending only on recordings their own team has gathered over time.

“Researchers in India already use global deep learning models for acoustic analysis, with individual modifications and data inputs to fine tune the model’s accuracy in the habitat they’re working in. But this is all happening in individual silos, wherein the data and resources are fragmented. With a one stop for a vast amount of data, all those efforts can be pooled together,” Khanwilkar said.

Listening to changing forest soundscapes in the rainforests of Western Ghats using passive acoustic recorders.

The IEN grew out of Project Dhvani, a collective that Khanwilkar, Choksi and avian ecologist Vijay Ramesh began as PhD students at Columbia University in 2018. Relaunched as the network in 2024, it drew over 600 people interested in ecoacoustics, formal expertise not required.

In 2025, a core group of ecologists and engineers put out the call, with a few ground rules: the recordings would be free for anyone to use under a Creative Commons licence, labelled and published, and anyone who sent in at least 30 minutes would get their name on the paper as a co-author. The finished paper credits 59 contributors.

“The bottleneck is that there is not enough training data from actual ecosystems in India. What we want is noisy data. We want that bird song, overlapping and competing, with all the insect buzz and human noise in the background, interrupted by rain and wind, rocks and leaves,” Choksi said.

“More representative data will make the AI models better at identifying species,” said Sonal Teotia, a Bangalore-based birdsong enthusiast who works in AI training and sent in her own recordings of birds and frogs.

Teotia’s husband, Sameer Singh, an electronics engineer, handled the incoming recordings and checked their quality, throwing out any that arrived without a species being named. Before the dataset went to Scientific Data, independent experts spot-checked a random 2 per cent, and 98 per cent held up. Most of it came from Kerala and Madhya Pradesh. Of the nearly 6,000 minutes, 3,311 are pinned down to the exact second and pitch of each call; the other 2,504 loosely confirm a species’ presence.

The data could hint at shifting migration patterns and changing behaviour. “If there’s a drought, we could even hear differences in the ways species call when the rivers are running dry,” Choksi said.

The collective, meanwhile, has hit its storage limit on Zenodo, a CERN Data Centre-backed research data repository, and the voluntary effort needs more hands for annotation and verification. But Khanwilkar, who is now scouting for funding, said: “The dataset paper is only the first step, we will continue to expand the data.”