Indic-Transcribe Review: Does IIT Madras’ New AI Model Really Beat Gemini at Indian Languages?

Every Indian tech outlet ran the same headline this week: IIT Madras has built an AI model that understands 26 Indian languages. What most of them didn’t do is check the one number that actually matters — the claim that this model, called Indic-Transcribe, beats Google’s Gemini 3 Pro at transcribing Indian speech. We did. Here’s what holds up, what’s unverified, and what it means if you’re building anything that needs to listen to India.

What Is Indic-Transcribe, Exactly?

Indic-Transcribe is a speech recognition model built by BodhanAI — an IIT Madras-incubated initiative that runs India’s national Centre of Excellence for AI in Education under the Ministry of Education — in partnership with AI4Bharat, the IIT Madras research lab behind most of the open Indic-language datasets used across the industry.

The headline numbers: 1.2 billion parameters, trained to handle 26 Indian languages plus English, and built specifically to deal with the messiness of real Indian speech — mixed-language sentences, regional accents, and non-standard scripts that most global models stumble on.

Professor Mitesh Khapra, who leads AI4Bharat and co-announced the model on X on August 17, described India as “a voice-first nation” — a fair framing, given how much of the country still prefers speaking over typing, especially outside metro cities.

The model goes live for public access on the BodhanAI platform on September 5, 2026. It isn’t available to test yet, which is worth remembering every time you read a benchmark claim about it right now — including the ones in this article.

The Benchmark Claim, Fact-Checked

This is the number every outlet has repeated without context: BodhanAI says Indic-Transcribe’s “Core” version scored a word error rate (WER) of 8.9, beating Sarvam AI’s Saaras V3 (10.1) and comprehensively outscoring Google Gemini 3 Pro, which came in at 18.9. The lighter “Flex” version scored 11.1.

For readers unfamiliar with the metric: word error rate measures how many words a speech model gets wrong out of every 100 — insertions, deletions, and substitutions combined. Lower is better. A WER gap of roughly 8 to 10 points, if accurate, is a genuinely large difference in real-world usability, not a rounding error.

Here’s the catch, and it’s an important one: these numbers come from BodhanAI’s own internal comparison, not an independent, third-party benchmark. No neutral evaluation — not from Hugging Face’s Open ASR Leaderboard, not from an academic paper, not from any outside lab — has verified this yet. Self-reported benchmarks from a team launching their own product are not automatically wrong, but they are not the same thing as a peer-reviewed result either. Until the model is actually public on September 5 and developers can run their own test sets against it, treat this as a strong claim worth watching, not a settled fact.

Core vs. Flex: Which Version Actually Matters for You

BodhanAI is shipping two versions, and the difference matters more than most coverage has explained:

The model’s stated range is also genuinely wide: BodhanAI says it can handle everything from Sanskrit shlokas and railway announcements to multilingual cricket commentary, and it covers lower-resource languages like Bodo (Assam) and Santali (Jharkhand) that are routinely left out of mainstream Indian AI tools — Bengali, Odia, Bhojpuri, Punjabi, Nepali, Haryanvi, Maithili, and Hindi are also on the list.

The Children’s Speech Pilot — What’s Confirmed vs. What Isn’t

One detail circulating in early social coverage claims the model was trained on over a million hours of speech and used a formal “child assent” and regulatory-clearance process. We couldn’t verify either claim against any primary source.

What BodhanAI has actually confirmed: pilots are currently running across 12 languages, using 12,000 hours of speech collected from students in grades 1 to 10, with parental consent obtained. That’s a meaningfully smaller and more specific claim than the inflated version making the rounds — and, frankly, still a solid foundation for education-sector use, given that children’s speech is notoriously hard for ASR models to handle well.

If Indic-Transcribe performs reliably on children’s voices at launch, that’s arguably a bigger deal for rural and regional-language edtech than the headline benchmark — most voice-based learning tools still fail badly on young or non-standard-accented speakers.

Who Should Actually Care About This

The Gap Nobody’s Talking About: Indian Sign Language

Here’s an angle almost entirely missing from the coverage so far. India is investing heavily in AI that understands 26 spoken languages and their accents — but Indian Sign Language, the primary mode of communication for millions of deaf and hard-of-hearing Indians, has no comparable national foundation model or funded initiative. What exists today is scattered: small academic research projects, individual college papers, isolated computer-vision experiments — nothing at the scale or institutional backing of Indic-Transcribe.

That’s not a criticism of BodhanAI specifically — voice is a genuinely harder, more commercially obvious starting point. But if “voice-first India” is the mission, sign language deserves to be next on the sovereignty roadmap, not an afterthought.

What to Watch Before September 5

Bottom Line

Indic-Transcribe is a real, credibly-backed model from a serious institution, not vapourware — the language coverage, the education-focused design, and the transparency around what’s still a pilot (the children’s-speech data, for instance) all point to a genuinely useful project. But the “beats Gemini” headline every outlet is running deserves an asterisk until someone outside BodhanAI runs the numbers. Bookmark September 5 — that’s when we’ll actually know.

FAQs

aitechnews.in will follow up with independent testing once Indic-Transcribe goes live on September 5, 2026.