वन धन · संताली भाषा सेतु
Van Dhan Santhali
A Hindi to Santhali bridge for Van Dhan MSP and scheme content, running entirely on Bhashini, with an in-app Ol Chiki to Devanagari converter and an honest readiness panel.
- Languages
- Hindi · Santhali in Ol Chiki script and Devanagari speech
- Channels
- Browser text translation · Browser Santhali speech playback
The problem
Tribal forest-produce collectors working through Van Dhan Vikas Kendras frequently speak Santhali. Minimum support price and scheme information reaches them in Hindi, which means it reaches the literate intermediary rather than the collector.
Sovereign language infrastructure exists for exactly this: Bhashini, the Government of India's language stack. But its coverage for low-resource tribal languages is uneven in a way that is not visible from a feature list, and a deployment that assumes uniform support fails silently in the field.
There is also a script problem hiding inside the language problem. Santhali is written in Ol Chiki, and the parts of the pipeline do not agree on it.
What AiSewak built
This deployment translates Van Dhan MSP and scheme content from Hindi into Santhali and speaks it aloud, running entirely on Bhashini. Translation uses AI4Bharat IndicTrans2 through the ULCA pipeline; the spoken output uses the IITM Santali text-to-speech service.
Between those two services sits the actual engineering. Bhashini's translation emits Santali in Ol Chiki, while the IITM speech service requires Devanagari input. Bhashini offers no Ol Chiki to Devanagari transliteration, and Aksharamukha does not support Ol Chiki either, so the conversion is implemented in the application.
That conversion is not a character swap. Ol Chiki is a true alphabet in which every vowel is a full letter with no inherent vowel, while Devanagari is an abugida where consonants carry an inherent vowel and need matras and viramas. The converter inserts them, and was validated end to end against the live speech service rather than against a specification.
The deliberate simplification is documented in the code rather than hidden: the modifier signs marking nasalisation and vowel length are dropped as a best-effort phonemic mapping, with the upgrade path named for when pronunciation fidelity has to improve.
The page also carries a readiness panel stating plainly what Bhashini deploys for tribal languages today. Santhali has translation and speech output but no speech recognition, so a full spoken exchange currently uses Hindi for voice input. That panel is the deliverable as much as the translation is.
How it runs
Language stack
A server-side Bhashini ULCA client running the standard two-step flow: a pipeline configuration call, cached for a day, followed by a compute call per request. No GPU, no third-party voice vendor, no data leaving the sovereign stack.
Script bridge
An in-application Ol Chiki to Devanagari transliterator that models the alphabet-to-abugida conversion, including consonant clusters, independent versus matra vowel forms, the anusvara, Ol Chiki digits and sentence punctuation.
Honest readiness
A published per-language table showing which of speech recognition, speech synthesis and translation Bhashini actually deploys, so a department planning a tribal-language rollout can see the gap before committing to a design that depends on it.
Stated limits
Machine translation and speech synthesis of a low-resource language can be imperfect, and the page says so: critical figures must be validated before real use. The demonstration is AI-disclosed throughout.
What a department can replicate
- 1
For any ministry or state planning a tribal-language service, the first reusable step is the readiness audit: establish, per language and per task, what the sovereign stack actually deploys today, and publish it. Most designs that fail in the field failed at this step, on paper, months earlier.
- 2
The second is to expect a script mismatch between pipeline stages and to budget for bridging it. Two components that both claim to support a language can still disagree about which script it is written in, and nothing in either API will say so.
- 3
The third is to build the honest degraded path rather than waiting for full coverage. Translation and speech output in the target language, with speech input in a widely spoken second language, delivers value now and upgrades cleanly when the missing model is deployed.
Scope the same pilot for your department
Two districts, a bounded knowledge base, a written list of what the agent will refuse to do, and a measurement step in the call itself. That is the shape every pilot above started from.