Community-owned speech data for endangered languages, licensed on their terms
TL;DR Oralbank trains speakers of endangered Indian languages to record and transcribe their own speech, then licenses the cleaned data to government programmes, universities and AI labs. The community keeps a say over every use and a share of every fee. It sells the slow groundwork that every language tool depends on.
India's 2011 Census recorded 2,843 mother tongues, and the central government's endangered-language scheme has so far flagged 117 languages as at risk, most with under 10,000 speakers[1]. Toda, spoken in the Nilgiris, is down to a few thousand people[1].
For most of these languages there's almost no digital text, so nobody can build translation or speech tools for them.
The elders who hold the songs, place names and plant words are getting older, and the recordings that do exist often sit with outside researchers and not with the community that made them.
Regulatory scope: no sector licence applies. Voice recordings are personal data under the DPDP Act, so every speaker gives recorded consent.
The seed pitched AI-powered immersive preservation, and I reframed it. No model can revive a language without hours of transcribed speech, and that data is the scarce part, so Oralbank sells the data work itself.
It recruits and pays local young people as recorders, trains them on a phone app that works offline, and runs structured sessions with elders - word lists, stories, everyday conversation. A second paid group transcribes and translates into Hindi or the state language. Machine help is used only for segmenting audio and spotting gaps.
Each corpus sits under a licence the community council signs off. It's open for schools and the community, paid and limited for commercial AI use, and a share of every fee goes back to a community fund.
Speakers and community groups use it and pay nothing. Funders pay per corpus. The Ministry of Tribal Affairs has funded single projects of ₹58.70 lakh and ₹3.122 Cr for tribal-language documentation and AI translation[1]. State tribal research institutes, university linguistics departments and philanthropies are the other buyers. AI labs that need low-resource speech data are a second, later payer for commercial licences.
The Sanchika portal for endangered-language audio went live in July 2025, and the scheme says it wants to document about 500 lesser-known languages[1]. A big tech firm funds Project Vaani to collect speech across India's districts[2], and dataset firms now list global AI labs among their clients[3]. Demand for this data is growing faster than the supply of trained community recorders.
There's no public figure for an endangered-language documentation market in India. The firmest signal is grant size, with ₹58.70 lakh and ₹3.122 Cr going to single tribal-language projects[1].
Next door, Vaani has 31,278 hours of audio across 105 languages but only 2,122 hours transcribed[2]. That tells you exactly where the bottleneck sits.
Project Vaani, run by IISc and ARTPARK, collects district-level speech and releases it openly under CC-BY-4.0[2]. One Indian data company builds paid speech datasets in the 22 scheduled languages for big global tech clients[3]. A Māori community media group in New Zealand holds Māori recordings under its own Kaitiakitanga licence[4].
Nobody in India focuses on the smallest languages with a community-controlled licence and revenue share.
| Difficulty | Moderate (36/100) |
| Starting capital | ₹40–80 L |
| Time to launch | 4–6 months |
The tech is light. The work is field trust and quality, and you need linguists who can design elicitation and check transcription in scripts that may not be standardised. Grant cycles are slow and each language is a small project, so the team has to run lots of them in parallel to stay solvent.
No sector licence applies. The obligations come from consent, copyright, field access and grant terms.
Building this outside India? The categories above usually have a local equivalent, so check with your own regulator.
Don't trust AI to transcribe or translate an endangered language on its own. The models have almost no data, so speakers must check every line.
Want to work out the AI layer for Oralbank? Talk to Yash →
Here's what can go south -
I have a soft spot for this one, but it suits a founder with a linguistics or field-research background and patience for grant work. If you expect software margins then it is the wrong fit. Prove first that one funder will pay a fixed price per transcribed hour and that one community will sign the licence you propose.
Plan for roughly ₹40–80 L, mostly on paying recorders and transcribers, linguists and field travel.
It can cover its costs through grants, but revenue is lumpy. Single government projects have been funded at ₹58.70 lakh and ₹3.122 Cr, and AI lab licences are a later bet.
Yes. A Māori community media group in New Zealand shows the community-licence model works. Check local data protection, copyright and indigenous-rights rules.
The free Oralbank validation kit has the line-by-line capital breakdown, the effort scoring, the consent and data-protection checklist, a 14-day validation sprint with pass/fail targets, a 90-day launch plan, 10 interview questions for funders and community leaders, and 5 assumptions to test. Download the kit →
Thinking about building Oralbank or something close to it? The make-or-break work is pricing a transcribed hour so that recorders get paid fairly and funders still sign. If you need a hand, you know where to find me. Talk to Yash →
For illustration only. Names are fictional and ideas hypothetical; sourced facts are real but may be out of date. Nothing here is legal, financial or professional advice, and Yash Khatri is not personally liable for any use of it. Read the full disclaimer.