When Manisha Warghade sat down to type her own language into a phone, the words weren’t there.
She is an associate professor in Navapur, a town in Maharashtra’s Nandurbar district, and a native speaker of Bhilali, one of the tongues that shade into one another across the hills where Maharashtra, Gujarat and Madhya Pradesh meet. She has worked on tribal languages for years. Yet typing one sentence in her own took a long time, she says, because the words were not available to type. The language had no settled place inside the phone, no keyboard that expected it.
That absence is what a tribal district set out to fill, under a name that aimed higher than a chatbot. In the district’s records the project ends in two words: Astitva Prakalp, the Existence Project.
The language at its centre is Dehwali Bhili, spoken by some 10 million people across western and central India and written down by almost no one. The plan was to gather enough of it as speech to train a machine to recognize it, translate it, say it back and then wire that machine into MahaVISTAAR, the Maharashtra government’s AI farm-advisory, which carries university-vetted crop and dosage advice into a phone.
At the end of it, a Bhili-speaking woman on a basic phone could call a number, ask about her crop or a sick child, and be understood by the state. Chief minister Devendra Fadnavis announced the first model as a first of its kind—a tribal language taught to an artificial intelligence machine.
The people who built it were after the larger meaning of the name: the identity and the undocumented knowledge a language carries, the words for a tree, a grain-store and a marriage rite that live nowhere but in Bhili. Saving the language, they kept saying in one room after another, meant saving the culture inside it. And to do even the first part, the district had to choose—one dialect out of at least four spoken within its own borders, and one use case, agriculture, out of the five the project began with.
No choice between dialects is small. “There’s a popular saying in linguistics that a language is just a dialect with an army,” says Manu Chopra, who runs Karya, a data company involved in the project. “Each dialect is a unique subculture.”
India is preparing to teach AI machines hundreds of such languages. And Nandurbar is where the method is being worked out first.
The door
The person who decided Warghade’s language should exist inside a machine is the district’s collector, Dr Mittali Sethi. She traces the idea not to a machine but to a door.
She is Punjabi, a dentist by training. She joined the civil service and was posted into tribal Maharashtra. Years earlier, in Melghat, she had picked up a little Korku, another tribal language, and was in a village following a mundubaba, a local healer. Here, she tried a sentence: “Tumha naam kaya?” (what is your name?)
Villagers answered, opened the door to a house and asked her inside. She sat on the floor with them. She had expected to be kept out. What let her in was not fluent Korku but the visible fact that she had tried. An imperfect sentence, she learned, opens a door a perfect one leaves shut.
The lesson hardened into a conviction during the pandemic, when governance turned out to rest entirely on communication. She had already proved it once: wanting breastfeeding tutorials that existed only in Marathi to reach Korku mothers, she built a makeshift recording booth in her one-room office.
Korku became the first language in the country to get those tutorials. Language, she decided then, was how governance reached people, or failed to. By the time ChatGPT arrived, the stakes were plain: Marathi was entering the new AI systems because it had books, users, subtitles—years of accumulated text. Bhili had none of that, and the absence was historical, not natural. She once messaged someone at Google to ask whether they could simply put Bhili in. A childish question, she calls it; the answer was data.
The radio station
When the work began it stopped almost at once, on a problem no one had planned for: the recording needed a room quiet enough for a machine to learn cleanly from a voice, and the district did not seem to have one. A team had come from Delhi, Bengaluru and Ranchi, and for two days they went building to building and found nothing. Then someone pointed them to a radio station.
It belonged to Jaywant Kapade, who runs Vikas Bharati 90.8 FM, the first community radio station in the district, built to carry crop and nutrition advice to villages. Kapade had never heard Dehwali Bhili until the morning they came to record it. He had spent a year at a film production house in Mumbai, long enough to know what a clean recording is, and when the team asked about acoustics he did not promise—he cut the same lines three ways: raw, middling, broadcast-clean—and let them choose with their own ears.
They chose the clean one. A studio built to broadcast a tribal language became the room where that language was fed into a machine.
What Phaito holds
The people who fed it were, for the most part, the district’s own agriculture officers—men who had carried the state’s advice into Bhil villages for years in Marathi, though nearly all of them are Bhili speakers themselves. They gathered one morning in the block agriculture office at Navapur, sat in a circle of plastic chairs under three turning fans, and were asked not about crops but why they did this.
Ravi Shankar Parve, the taluka officer who had made the project his own, answered with the clock: in 15 or 20 years, Dehwali Bhili had drifted under schooling. Hindi words had slipped in unnoticed. A language with no script, if not set down now, will disappear with the old speakers.
Tulshidas Wadvi, a sub-divisional officer who calls himself not an agriculture officer but a worker for the Dehali Bhili language, cited words that could not be carried into any other language. Pamro, a kind of tree with no name in the bigger tongues. Phaito, the platform of lashed bamboo on which grain was stored before there were sacks—a whole vanished method folded into one word. Hikam, the storage of raw mangoes through the heat so they ripen slowly, to be eaten later with the grandchildren—not just a name, he said, but a story.
The fear in the room was that the machine would hold the wrong thing. People rushing a deadline mix Marathi or Gujarati into the work, one officer warned, and the model cannot tell the difference. Train a distorted Bhili in and it cannot easily be undone. They knew this because it had already happened. The first round of data came back thin and heavy with Marathi; the second, done in even less time, took on Gujarati from the border villages as also Hindi. The model had not been released. Its testing was still ongoing.
Three hundred kilometres away, the engineer who orchestrated the technical work explained why this is so hard.
Jagadish Babu, the chief operating officer of Ekstep Foundation, a non-profit, had been at Tembhli, a village where the state issued India’s first Aadhaar number to a tribal woman, in 2010.
He frames the data problem as voltage: wire many half-volt batteries in parallel and the diode still will not glow; only past a threshold, around 1.2 volts, does it light. For an oral language that means a one-time effort of 1,000 to 1,500 hours of quality speech, and a real use waiting at the other end.
“If there is a landing strip, the plane aligns itself,” he says. “If there is no landing strip, where will the plane go?”
By the district's own count, the first phase gathered about 68 hours of speech—60 spontaneous, six in the studio, two conversational—along with 40,000 transcribed sentences, 25,000 of them on agriculture. It is well short of the destination.
The file not taken
The reason the room filled at all was a poet named Kamlesh Mahale, who entered the project through a Facebook advertisement as Karya (the data company)’s regional coordinator in Nandurbar. He turned out to be the one person who could make it move.
As a young man, he found his way to an older tribal-rights writer on whose shelves he first read Camus and Sartre. “I had a lot of privilege,” he says, “even while living in Nandurbar.”
He did not recruit through any system; he recruited through two decades of knowing who in the district could be trusted with a word. If you are from the region, he says, people believe in your honesty, and that trust was the thing the project ran on.
He was also honest about its limits. He had wanted vans to reach the deep villages past Dhadgaon where the oldest, least-mixed Bhili survives, and the budget had said no. The purest speakers were the hardest to reach, and the easiest data, from the young and connected, were the most mixed.
His part, he said, was only the collection. What they would do next, he did not know. The project had begun with five subjects—agriculture, education, administration, culture, and one more—and narrowed to agriculture alone for the deadline.
Without culture, how can a language be whole?
Outside the file sat the wedding songs, the harvest songs, the masked dances that carry tribal freedom fighter Birsa Munda’s story, and a protest bhajan the officers sang in the room—a thief pleading before a judge that he is a devotee of Shiva. None of it can go into an agriculture chatbot.
People and platform
The machine itself was built by a chain of people. Karya wrote a first modest proposal for text alone. Parimal Singh, the project director of PoCRA, the state’s climate-resilience agriculture programme under which MahaVISTAAR was built, Dr Mittali Sethi and Amitabh Naksar of Bhashini, the Indian government’s AI-powered language translation platform, refused it, insisting on the whole deal: speech recognition, translation, text-to-speech.
Then EkStep Foundation came in. The non-profit was already the voice behind MahaVISTAAR and Singh got them to Nandurbar once collection was already under way.
EkStep had voice engineers and they were clear about where their expertise stopped: for now, the project was a data problem more than a voice one. So they kept their hands off the data.
The company that gathered the data would not also build the model; that went to AI4Bharat, the language-AI lab at IIT Madras. The model and its data would be open by default, lodged in AIKosh, the national AI library the government had opened a year earlier, so the next person chasing a forgotten language would not start from zero. Bhashini would host it.
And EkStep would not let them grade it on a soft curve; it made them run the same tests the world runs on any voice model, then had the method written down and taught.
The cost tripled.
When the chief minister’s launch was fixed for February this year, Dr Sethi did the arithmetic backward and found she had a month; the team did not sleep. Contributors were paid by the word—a rupee to record, more to transcribe and validate. Of the roughly ₹27 lakh spent in that phase, ₹18 lakh went back to the people who gave their voices. Building a language is labour, she says, and the person who picks up a phone and gives an hour of it should be paid.
The first model, shown that day, did not work—answers came back in five, seven, ten seconds. Then they found the deeper flaw: the model had been trained on clean studio audio, but the farmers it was for would call from feature phones over eight-kilohertz telephone lines, rougher and lower, and a model built on the clean version fails on the real one.
They did not settle for the slow version, and the reason was the people it was built for. “A demanding customer,” Jagadish Babu says, “is the primary requisite for a good quality language model.”
So the second version was re-collected over actual phone lines. Because of the launch timing, IIT Madras built the model and Karya built a parallel one as a backup. It is built now, hosted, and is being tested with the state agriculture department.
The work has since passed from the district to the state, with the validated dataset placed as an open repository on Bhashini and ownership resting with the tribal-welfare department.
The district’s roadmap names what comes next: three more dialects—Mathwadi Bhili, Mavachi and Pawari—and the domains held back for the deadline, among them education, administration, and a bot for the health workers who reach the villages.
State view
The district has been here before. A generation ago, the rising water of a great dam drowned villages whose people were these same Bhils. Their displacement was logged as development. In 2010, the state came to Tembhli and gave a tribal woman the country’s first Aadhaar number, counting her by reducing her to twelve digits. Now it has come a third time, with a microphone and a model, after the language has been narrowed to one dialect and scoped to the price of seed. Each arrival has let the state see these people a little more clearly.
What the machine cannot yet hold, a singer keeps. Amrut Nathu Padvi performs the Bhili songs—of Birsa Munda, of the forest, of the long struggle with the state. During festivals, the performances run so long, he says, there is no time even to take a phone call. He does not write the songs; others do. His part is to sing them.
Take the songs and the dance out of the language and what is left? He did not answer in the language of policy. They will close the temple, he says. Then the king cannot come.