Voice AI in Punjabi and Hindi — what's actually possible in 2026

A lot of AI writing pretends every language is the same problem. It isn't. English voice models have had a decade's head start — oceans of training data, armies of people fixing edge cases. Punjabi and Hindi have had a fraction of that. The codemix people actually speak (half a sentence in one language, half in another) has had almost none.

So when a client asks "does voice AI work in Punjabi?", the honest answer in 2026 is: mostly, for the first time. Anyone selling you Punjabi voice AI in 2024 was overselling. Today they'd be mostly right — but ask them which words their model still stumbles on. That question is the whole article.

Where we are in 2026

The last twelve months changed the ground under our feet. Native-audio voice models — Gemini Live, Sarvam's Bulbul line, OpenAI's realtime models — stopped treating speech as "transcribe, think, synthesise" and started listening and speaking directly. That one architectural shift did two things that matter for Indian languages.

First, latency collapsed. A round trip that used to take two to three seconds — long enough for a caller to say "hello? hello?" — now lands under 800 milliseconds on a decent connection.

Second, and less obvious: with no separate speech-to-text step, the model stops mangling a Punjabi sentence into the wrong English words before it even starts thinking. It hears the sound and answers the sound. For a language full of tones and retroflex consonants that English STT routinely fumbles, that's a real jump.

What actually works

Here is what we now consider reliable enough to put in front of a paying customer's customers:

• Natural conversation in Punjabi, Hindi, or English — with no language toggle. The caller speaks whatever they speak; the model follows.
• Code-mix. "Mujhe do chai chahiye" or "appointment book karni si" — the model handles the switch mid-sentence without flinching. This was science fiction two years ago.
• Domain vocabulary — product names, service types, locality names — once the model is primed with the business's own words. A dental clinic's "scaling and polishing" or a boutique's "phulkari dupatta" lands correctly.
• Interruption handling. The customer cuts in halfway through the AI's sentence; the AI stops, listens, and adapts. Real conversations are full of interruptions, and this is what makes it feel like talking to a person rather than a menu.
• Sub-second latency on 4G in urban India — the difference between "this feels alive" and "is this thing broken?"

What still trips it up

You'd rather hear the limitations from us than find them in a live call. In 2026 the model still stumbles on:

• Deep regional dialects. Malwai and Doabi Punjabi, thick rural Haryanvi-inflected Hindi — the model drifts, guesses, sometimes asks the caller to repeat. It's not lost. It's just not fluent the way it is with "standard" registers.
• Loud, messy environments. A workshop floor, a construction site, a busy kitchen — background noise confuses the voice-activity detection, and the AI either talks over the caller or waits too long.
• Very elderly speakers with strong accents and slower, softer speech — recognition accuracy dips, and patience on both sides runs out faster.
• Slang with a six-month half-life. The words teenagers use this year the model learned about last year. Fine for a clinic; risky for a youth-facing brand.
• Long numbers spoken fast. A ten-digit phone number rattled off in one breath will occasionally drop a digit. We design around it — read it back, confirm — instead of pretending it's solved.

The stack we actually use

No mystique here. For most production voice work today we run:

• Gemini Live (gemini-2.5-flash-native-audio-preview-12-2025) for the voice-native turns — it carries the conversation in the caller's language.
• Sarvam STT/TTS as a fallback and for text-only channels where native audio is overkill.
• Claude Haiku 4.5 for reasoning and tool calls — checking availability, writing to a booking system, deciding when to hand off.
• Voice-activity detection tuned per environment — a quiet clinic and a noisy showroom get different sensitivity settings. This single knob decides whether the experience feels calm or chaotic.
• Rate limits and budget caps in production — because a voice model left uncapped is a way to wake up to a large bill.

Try it

The voice demo on our home page runs this exact stack. If you're visiting from Punjab, it defaults to Punjabi — tap the mic and talk. Two minutes, no signup. It's the fastest way to judge where the tech actually is, instead of taking our word for it.

What's next in 2026–27

Three things we're tracking. Emotional-tone detection is maturing — the model starting to hear that a caller is annoyed and drop its tone to match. Whisper-and-shout handling is getting better, so someone in a loud market or speaking softly in a waiting room both get understood. And on-device inference is inching forward, which would cut cloud dependency, drop latency again, and keep more data on the phone. None of it's finished. All of it's closer than a year ago.

One takeaway: the question has changed. It used to be "can voice AI even work in our languages?" Now it's "which conversations do we give it, and where do we keep a human?" Better question to be asking.