Hindi, Gujarati, and Hinglish voice agents: what actually works

Voice agents can now hold natural phone conversations in Hindi, Gujarati, and the Hinglish that most Indian callers actually speak. But “supports Hindi” on a feature list and “handles your customers’ calls in Hindi” are different claims. The gap between them is where most disappointing deployments come from.

This is what we have learned about making multilingual voice agents work on real Indian phone lines.

The phone is harder than chat

A chatbot reads clean text. A voice agent has to hear it first, and phone audio is difficult:

• Phone lines carry less sound. Standard phone calls are narrowband, which strips out much of the detail that speech recognition relies on. A model that works well on a studio recording can struggle on a mobile call from a moving auto.
• Callers do not choose a language first. In chat, the customer’s first message tells you the language. On the phone, they just start talking, and they may switch three times in one sentence.
• There is no backspace. If the agent mishears a name or a date, it has to notice and recover in the conversation, not in a form.

Any vendor demo recorded in a quiet room tells you very little about how an agent will do on your callers’ calls.

Code-mixing is the norm

Very few Indian callers speak “pure” Hindi or “pure” English on the phone. A typical booking call sounds like this:

“Haan, mujhe kal ka appointment reschedule karna hai, Thursday evening ho jayega?”

That sentence is Hindi grammar with English nouns and verbs, and it is completely normal. A voice agent that only handles one language at a time will either misunderstand it or answer in stiff, formal Hindi that sounds nothing like the caller. Neither builds confidence.

The agent needs to understand mixed speech and reply in the same register. If the caller speaks Hinglish, the agent should too.

Names, numbers, and addresses are the hard part

Understanding intent, such as “I want to reschedule”, is the easy part. Capturing details accurately is harder:

• Indian names have many spellings, and speech models often guess the most common one.
• Phone numbers are spoken in groups, in English or in Hindi, sometimes both.
• Dates and times come in many forms: “parso”, “next Monday”, “do baje”, “2 o’clock”.
• Locality names are often not in the model’s vocabulary at all.

The fix is not a better model alone. It is good conversation design. The agent should read every important detail back to the caller (“So that’s Thursday, the 9th, at 6 pm with Dr Shah, is that right?”), and for anything that must be spelled exactly, send a WhatsApp message the caller can correct.

Let the caller lead the language

Do not start the call with a language menu. Open in the language most of your callers use, often a light Hinglish greeting, and let the agent follow the caller from there. If they answer in Gujarati, it continues in Gujarati. If they switch to English for a technical term, it carries on without missing a beat.

Keep the voice consistent while the language changes. Swapping to an obviously different voice when the language changes feels like being transferred, and callers notice.

Write the script in the language, do not translate it

A script written in English and translated into Hindi usually sounds like a government form. Real callers say “aapka appointment confirm ho gaya hai”, not a formal Hindi translation of “your appointment has been confirmed”.

Write the key phrases natively in each language, in the register your customers use. For most businesses that means polite but conversational: “aap” rather than “tum”, but not textbook Hindi. Get someone who speaks the way your customers do to read the script aloud before it goes live.

Test on your own calls, not a demo

The single most useful thing you can do before choosing a voice agent is to test it on recordings of your own customers. Pick twenty or thirty real calls that cover your common call types, your callers’ accents, and your typical line quality. Then check:

1. Did it understand what the caller wanted?
2. Did it capture names, numbers, and dates correctly, and read them back?
3. Did it reply in the caller’s language and register?
4. Did it transfer the calls it should have transferred?

This is how we confirm language coverage in every pilot, because it is the only test that predicts what will happen on your line.

Always keep a human path

No voice agent understands every caller every time. What matters is what happens when it does not. Set a simple rule: if the agent has failed to understand the caller twice, it apologises and transfers to a person, with the transcript attached so the caller does not have to start again. A caller who gets a fast handoff forgives the misunderstanding. One who is stuck in a loop does not.

A checklist before you go live

• The agent opens without a language menu and follows the caller’s language
• It handles mixed Hindi-English (and Gujarati-English, if relevant) in both directions
• Every name, number, date, and time is read back for confirmation
• Key phrases are written natively, not translated
• It has been tested on your own recorded calls, across your callers’ accents
• It transfers after two failed attempts, with the transcript attached
• Someone reviews a sample of transcripts every week after launch

Where to start

If your callers speak a mix of languages, start with the call types that are most predictable, such as bookings, timings, and status checks, and expand once the agent has proved itself on your lines. See how we build multilingual agents on our AI voice agent development page, read how clinics use them in our AI receptionist setup guide, or book a free call audit to have your own calls tested.