AI That Speaks Indian — Hindi, Gujarati, and the Rest

Part 8 of the series "AI, Without the Hype". Start at Part 1.

AI now works in Indian languages. Not perfectly, but usefully. And that single shift matters more than any clever English chatbot, because it is what will finally bring AI to the next half-billion Indians who never spoke to a computer in English in their lives.

For years, AI in India meant typing in English to a tool that answered in English. That was fine for the people who already had English. It did nothing for the shopkeeper in Rajkot, the farmer in Vidarbha, or my own grandmother. The technology was real. The reach was not. That is the gap closing now, and it is worth understanding why.

Why English-first AI quietly failed most Indians

Here is an uncomfortable number to sit with. The vast majority of Indians do not think in English. They speak Hindi, Gujarati, Tamil, Bengali, Marathi, and dozens more. English-first AI asked them to translate their own thoughts into a second language just to ask a question. Most people simply did not bother.

It failed quietly. Nobody filed a complaint. They just did not show up. A tool you cannot speak to in your mother tongue is not a tool for you — it is a tool for someone else that you are allowed to watch.

I saw this clearly while building SmartON, our assistive AI for the visually impaired, which now has 17,000+ users. A blind user in a small town does not want a beautiful English interface. He wants the phone to talk to him in the language his mother used. The moment we got that right, usage changed. Language was not a feature. It was the door.

The code-switching reality

Now the harder part. Indians do not speak one language at a time. We mix.

In a single sentence, a person in Ahmedabad might say something like, "Mને એ file મોકલી દે, urgent છે, by evening જોઈએ." That is Gujarati, English, and Hindi rhythm braided together in one breath. We do this naturally. We do not even notice.

This is called code-switching, and it is the thing most AI got wrong for the longest time. A system trained to expect "pure" Hindi or "pure" English breaks the instant you mix them. And we always mix them. Real Indian speech is not clean. It is a beautiful mess.

So the test of whether AI truly speaks Indian is not whether it handles textbook Hindi. It is whether it can keep up with a real human switching languages mid-sentence, the way we actually talk over chai.

What is genuinely good now

Let me be honest and specific, without hype.

A few Indian efforts have made real progress on language. Sarvam AI, a Bengaluru company, has been building models focused squarely on Indian languages, and in February 2026 released larger models aimed at reasoning in Indian contexts. Their earlier open work, OpenHathi, was about teaching Indian-language skills to existing models cheaply. Krutrim, from Ola, built a model that understands around 22 Indian languages and generates in about ten, including Gujarati. These are real, not vapour.

What is genuinely good today:

• Speech-to-text in major Indian languages is now usable for everyday tasks, not just demos.
• Translation between Indian languages and English is good enough for understanding, even if not for poetry.
• Voice assistants can hold a basic conversation in Hindi and a handful of regional languages.

What is still rough: deep reasoning in low-resource languages, sarcasm, regional slang, and the long tail of dialects. A tool might handle Ahmedabad Gujarati and stumble on Kathiawadi. We are early. But "early and useful" beats "absent."

For the engineering view of how these systems actually handle multiple Indian languages under the hood, I have written a more technical piece on multilingual AI for Indian languages.

What building MIRA taught me

MIRA is a multilingual voice AI router we built to handle exactly this — Gujarati, Hindi, and English code-switching in one conversation, routing each request to the right place.

Three things it taught me.

First, the model is the easy part. The hard part is the messiness around it: noisy audio, a user who starts in Hindi and ends in English, a name that exists in no dictionary. Real life does not arrive clean.

Second, you must design for switching, not just for languages. It is not "Hindi mode" and "English mode" with a toggle. It is one person, one breath, three languages. If your system needs the user to pick a language first, you have already lost them.

Third, accuracy is not the only thing that matters. Speed and dignity matter more. A reply that takes ten seconds in perfect Gujarati is worse than a fast, slightly imperfect one. People forgive small mistakes. They do not forgive being made to wait, or being made to feel stupid.

If you want the deeper technical walkthrough of how MIRA routes and handles all this, I have written about the MIRA AI router in detail.

Why vernacular AI is the real unlock for India

Step back and the picture is simple. India does not have an AI access problem because of cost or phones. Smartphones are everywhere and data is cheap. The real barrier was language.

When AI speaks your language, a whole population that was watching from outside can finally walk in. A weaver can ask about a government scheme in Marathi. A small trader can sort his accounts by speaking Gujarati. A student in a Tamil-medium school can learn from a tool that does not first demand she master English.

This is why I believe vernacular AI, not English chatbots, is the actual story of AI in India. English AI made headlines. Indian-language AI will make the difference. The first impresses investors. The second reaches Bharat.

AI did not arrive in India when it learned to code. It arrived when it learned to listen in our languages.

This is Part 8 of "AI, Without the Hype." Next, Part 9 — how a small business can actually use AI without wasting money.