India AI Digest — Thursday, June 18, 2026

• Gnani.ai launched Prisma v2.5, an India-hosted Indic speech-to-text model trained on a claimed 14 million hours of proprietary Indic speech across 12 languages, pitched at regulated BFSI, insurance and healthcare voice automation — with its accuracy edge over ElevenLabs, Sarvam and Microsoft self-reported, not yet independently benchmarked.

---

VOICE AI · INDIC LANGUAGE · BFSI · June 17, 2026

Gnani.ai launches Prisma v2.5, an Indic speech-to-text model built for telephony

Gnani.ai, a Bengaluru voice-AI company, said on June 17 it launched Prisma v2.5, a speech-to-text model it reports was trained on 14 million hours of proprietary Indic speech across 12 languages. The company positions the model for code-switched conversation, ambient noise and dialect variation on compressed telephony lines, hosts it in Indian data centres for low-latency real-time use, and targets BFSI, insurance, healthcare and compliance telephony. Gnani claims lower word- and character-error rates than ElevenLabs, Sarvam AI and Microsoft; those comparisons are the company's own, reported via Inc42, and are not independently benchmarked. No pricing was disclosed.

From the room. "Most ASR models are built for ideal studio conditions. Indian calls happen over compressed network lines, in at least two languages inside a single sentence, in accents no studio corpus has ever captured. Gnani Prisma v2.5 is built for that reality." — Ganesh Gopalan, Gnani.ai (company statement, reported by Inc42).

What this means. The defensible part of this launch is the training target, not the leaderboard claim. Telephony Indic speech — 8kHz compressed lines, two languages inside one sentence, call-centre ambient noise — is a distribution that studio-trained ASR systematically misses, and a model whose training distribution Gnani says bakes in that noise and code-switching — a corpus its contact-centre business is positioned to have collected — would be a real specialization rather than a re-skin. The source describes the 14M hours only as "proprietary Indic speech," so read the telephony-corpus composition as inference grounded in Gnani's business, not as disclosed fact. What is not yet checkable is the headline. The claim that Prisma v2.5 beats ElevenLabs, Sarvam and Microsoft on WER and CER rests on Gnani's own numbers from a single trade-press report; there is no published evaluation, no named test set, and no third-party reproduction. Read it as a vendor benchmark until one exists.

On the substance gradient, Gnani sits where it sat last month: a company that ships voice products but has not put a benchmarked model on the table. When the IndiaAI Mission named it to the indigenous foundation-model cohort, the open question was disclosure — whether the shipping would be matched by the kind of evaluation Sarvam and AI4Bharat publish as routine. Prisma v2.5 is a commercial launch with marketing-grade accuracy claims, not that disclosure. The product may well be the best-fit STT for Indian telephony; the materials released so far don't let a buyer confirm it without running their own bake-off.

The strategic lever, if the accuracy holds up, is procurement rather than capability. An India-hosted, low-latency STT tuned for regulated voice workloads removes a data-residency objection that BFSI and insurance compliance teams raise against US-hosted speech APIs. That is a narrower and more durable advantage than a benchmark line.

India angle. The buyers here are high-volume voice operations — bank and insurance call centres, collections, healthcare front-desks — where the actual purchase criteria are accuracy on degraded lines, latency, and where the audio is processed. Data residency is a gate, not a tiebreaker, for regulated BFSI and insurance telephony, and India-hosting clears it directly. The unit-economics read is the gap: Gnani frames this as a cost-and-latency play for high-volume automation, but with no pricing disclosed there is no rupee-per-minute figure to test against the incumbent cloud STT vendors a contact-centre would otherwise use. The case for Prisma is a procurement case until the price and an independent accuracy number both land.

Behind the news. On May 30 the IndiaAI Mission named Gnani.ai, alongside Soket AI and Gan.ai, to build indigenous foundation models next to Sarvam — Gnani's slot was a roughly 14-billion-parameter voice model (covered in the May 31 digest). That digest's forward signal was specific: whether any of the three would publish a foundation model with disclosed benchmark results rather than a launch post. Prisma v2.5 is the first ship from that cohort, and it lands closer to the launch-post end — a commercial STT with company-stated comparisons, distinct from the ~14B foundation model the selection promised. Gnani also carries a ₹177.27 crore IndiaAI allocation (per the April 4 digest), so the disclosure question is not academic; it is the standard the public-money cohort set for itself.

What to watch. An independent word/character-error-rate result for Prisma v2.5 on a named Indic ASR benchmark — or a named production deployment at a BFSI or insurance customer that cites India-hosting as the deciding factor. Either converts the current vendor claim into evidence. "Lower error rates than ElevenLabs" without a test set is not yet that.

See also: IndiaAI Mission compute crosses 34,333 GPUs; Soket, Gnani, Gan named · Sarvam turns unicorn with $234M Series B

Source: Inc42. June 17, 2026. →

Confidence: medium. Single tier-2 source (Inc42); the accuracy advantage over ElevenLabs, Sarvam and Microsoft is company-stated and not independently verified; no pricing disclosed.

---

Position movements