Offline voice Q&A in Indian languages on a $30-class device. No internet, no latency, no tokens.
Bhasha is the Indic voice family: speech-in, speech-out question answering in Indian languages, small enough to run fully offline on edge hardware in the $30 class. Built for schools, field services and citizen-facing programmes where connectivity, latency and data residency all argue against the cloud.
Offline Indic voice Q&A · vernacular text · edge devices
Classroom & learning assistants
Spoken Q&A on curriculum content in the student's own language, on a device that never needs a network.
Field-service voice
Hands-free, offline guidance for technicians and health workers in vernacular languages.
Citizen-service kiosks
Scheme eligibility, forms and instructions answered by voice at the point of service.
Vernacular document Q&A
Regional-language notices, circulars and forms made conversational.
Edge devices, fully offline — 4-bit quantised for $30-class hardware; also deployable on-premise.
POC phase; under evaluation for a state-scale school deployment.
References are anonymised by default; named references on request under NDA.
Bhasha 1.0 vs Google Gemma 4 — head-to-head on Indic voice tasks
Composite task accuracy on an Indic-language task suite: the questions and document types our clients actually pay for, not a generic leaderboard.
Internal evaluation, August 2026. Gemma 4 run with published weights, same prompts and hardware. Golden set per domain, blind human + AI-judge scoring. Methodology and raw runs available on request.
What comes with every Bhasha deployment
Provenance register
Base model, version and licence recorded for every deployment — the first artefact an auditor or a procurement team asks for.
Golden set per domain
Every family carries a curated evaluation set drawn from the document types its clients run on. Releases are gated on it.
Blind human + AI-judge scoring
Outputs scored by domain reviewers without knowing which model produced them, cross-checked by an AI judge. Raw runs available on request.
Frozen in production
The model that served your customers on Friday is the model serving them on Monday. No silent upgrades, no vendor deprecations, no pricing changes.
Related reading
Run Bhasha on your own documents.
Send 20–50 representative documents or transcripts — anonymised is fine — and the acceptance rules your team uses today. We run Bhasha 1.0 on infrastructure you control, score it against your criteria, and hand back the results and the harness. Two weeks, fixed scope, no production data.