Building a Hindi voice assistant that actually understands Indian parents

When we first built FamAI, it was text-only. You typed your question and got a response. Clean, simple, and completely wrong for the audience.

Indian parents in their 60s and 70s use WhatsApp primarily through voice notes. Typing is slow for them — smaller screens, changing eyesight, keyboards that weren't designed for Hindi transliteration. But they'll send a 45-second voice message without hesitation. They already do this dozens of times a day with family members.

Voice wasn't a feature we added. It was the feature we should have started with.

Voice notes > text input for this demographic

The practical reason is speed. Your parent can speak 10 words in 3 seconds. Typing those same words — especially if they're switching between Hindi and English, which most do — takes much longer and requires attention to autocorrect that's constantly wrong.

The deeper reason is naturalness. When your parent sends a voice note to their sibling saying "kal doctor ne kaha BP thoda upar hai, Amlodipine badh gayi hai", that sentence is effortless. Asking them to type the same thing creates friction that makes them either abbreviate awkwardly ("bp up doctor said") or give up.

FamAI processes voice notes identically to text. There's no separate "voice mode." Your parent records a voice note the way they always have, sends it, and gets a response. The abstraction is invisible.

Hinglish is the real language

Most Hindi voice assistants are built for one of two languages: formal Hindi or English. The reality is that educated urban Indians over 60 speak neither consistently. They speak Hinglish — a fluid mix of Hindi and English that code-switches mid-sentence without notice or pattern.

"Mujhe aaj doctor ke paas jaana hai — mera next appointment kab ka tha?" Half Hindi, half English, entirely natural. A system that only handles pure Hindi will catch the first part but trip on "appointment." A system optimized for English won't understand the context of "doctor ke paas jaana."

We trained FamAI to handle this code-switching, which required using examples of how Indian seniors actually speak — not textbook Hindi, not call-center English.

The dignity argument

There's a failure mode in voice assistants that we were explicitly trying to avoid: the system that makes you feel stupid.

If a voice assistant regularly fails to understand normal speech and asks you to repeat yourself, older users assume the problem is with them — not with the technology. They simplify their language: shorter sentences, fewer words, slower speech. The technology has effectively trained them to communicate worse. That's not acceptable for a product meant to help people.

Our design principle was that FamAI should do the work of understanding natural speech — not make parents do the work of simplifying it. Your parent shouldn't need to say "sugar 168" when they naturally say "aaj subah sugar check kiya, 168 tha, fasting tha." Both should work equally.

What we still can't do well

It's worth being honest about limitations.

Heavy background noise — a kitchen, a crowded room, a television — significantly degrades voice note quality and therefore our understanding of it. We handle quiet, clear recordings well. We don't handle poor audio well.

Very strong regional accents — Bhojpuri-inflected Hindi, certain South Indian English patterns — are areas where our accuracy is lower than we'd like. We're working on it, but we're not there yet.

When we don't understand a voice note, FamAI asks the parent to repeat or type instead. This is rare with clear audio, but it happens.