Sarvam AI began accepting applications this week from Chennai-area tier-two radiology cooperatives for grants that unlock a curated Tamil–English medical narrative corpus, provided hospitals contribute de-identified CT and ultrasound reports and agree to third-party bias audits before deploying Saaras speech-to-text in reading rooms, according to programme terms published on the Bengaluru lab’s site and interviews with two co-op administrators in Tambaram and Ambattur.
What the grants cover
Selected co-ops receive API credits for Saaras v3 streaming transcription, Bulbul text-to-speech for patient-facing summaries, and office hours with Sarvam’s clinical linguists. The corpus itself mixes open radiology phrase banks with co-op contributions—anonymized impression sections, measurement clauses, and Tamil code-switching patterns common when radiologists dictate for mixed-language staff.
Grants are not equity investments. Sarvam retains license to improve foundation models on aggregated patterns but contractually bars shipping identifiable patient strings back into public weights. Co-ops keep ownership of raw exports stored in India-based object storage with keys held locally—a requirement several administrators said Ayushman Bharat Digital Mission-aligned vendors already meet.
Why tier-two radiology co-ops
Chennai’s peripheral hospitals share MRI and CT slots across member clinics, producing high report volume but thin IT staff. Manual transcription to Tamil for referring physicians in Kanchipuram and Chengalpattu districts often lags same-day reporting targets. Sarvam’s HealthPlix partnership showed that Indian-accent clinical speech needs domain-tuned models; radiology co-ops present a narrower vocabulary than general outpatient consults but demand precision on millimeter measurements and contrast phases.
Co-ops must demonstrate at least six months of digital report archives and a named data protection officer before grants activate—filters meant to exclude single-clinic boutiques that cannot sustain audit trails.
Evaluation and safety gates
Sarvam requires pre-deployment word-error-rate benchmarks on held-out Tamil dictation samples and prohibits auto-release of reports without radiologist sign-off. Models may suggest structured fields—organ, lesion size, BI-RADS categories—but final PDFs must pass human review. Emergency keyword lists flag phrases suggesting pneumothorax or intracranial bleed for priority queues.
MeitY observers have not certified the programme; hospitals remain responsible for Clinical Establishments Act compliance. Sarvam’s docs note Saarika v2.5 deprecation and steer new clinical work toward Saaras v3 APIs—a migration co-op IT leads said will require retesting integrations built during pilot planning last quarter.
Technical stack in practice
Reading rooms route dictation microphones through on-prem edge boxes that stream audio to Sarvam endpoints; transcripts return over TLS to radiology information systems via HL7 ORU messages in early integrations. Bulbul voices read simplified Tamil summaries for patients waiting in crowded outpatient halls—a use case ethicists asked co-ops to limit to non-diagnostic explanations unless counselors are present.
Data contribution economics
Co-ops that contribute more than ten thousand de-identified reports receive additional inference credits and priority support queues. Smaller members can participate as consumers only, paying standard API rates after grant periods expire. Sarvam declined to disclose grant totals but said funding comes from its research budget rather than venture round tranches—a distinction startups lawyers said matters for hospitals worried about vendor lock-in from subsidized pilots.
Competitive landscape
Google and Microsoft cloud speech APIs support Tamil, but radiology administrators in pilot calls said generic models mangled drug brand names and anatomy terms anglicized in local practice. Sarvam’s pitch is Indic-first training plus on-shore processing commitments. Legacy transcription bureaus in Chennai warned that automation could squeeze manual typist jobs; co-ops responded that backlog clearance is the nearer pain point.
What regulators will watch
State health departments may ask whether speech drafts become medico-legal records before sign-off. Sarvam’s contracts place liability on signing radiologists, mirroring assistive-coding tools in the United States. Tamil Nadu’s e-health procurement cells have not issued tenders tied to the grant, leaving adoption voluntary for now.
Next milestones
First cohort results are due by December, when Sarvam plans a public technical note on Tamil radiology WER improvements versus baseline Saaras. Co-ops that miss audit deadlines lose API credits—a stick intended to keep de-identification pipelines honest as Chennai’s tier-two imaging market races to clear post-monsoon scan backlogs.
