Hinglish voice agents: how to handle callers who mix Hindi and English

"Haan, mujhe Tuesday ka appointment chahiye, morning mein." That sentence has three languages' worth of grammar and vocabulary in it, and it is how a lot of India actually talks on the phone. A voice agent that only handles pure Hindi or pure English will mishear it, answer stiffly, or both.

Listening: choose the right transcriber

Speech-to-text is where code-mixing breaks first. A model tuned only for English will spell Hindi words phonetically (or drop them); one tuned only for Hindi may do the same to English nouns like "appointment", "order" and "delivery".

What works:

• Pick a transcriber that supports Indian languages and code-mixed speech, and set the agent's language to Hindi (hi) rather than English for Hinglish callers.
• Add keyterms for the words your callers will say that a general model may not know: your brand, product names, doctors' names, localities.
• Keep a fallback transcriber configured, so a provider outage doesn't drop calls.

Thinking: tell the model how to reply

The language model will mirror the caller if you ask it to. Add these lines to the prompt:

• Reply in the same mix of Hindi and English the caller uses.
• Use common English words people say in Hindi sentences (appointment, order, payment, delivery) rather than formal Hindi equivalents.
• Write Hindi in Devanagari in your replies, with English words in Latin script.

The last line matters for the voice: most Indian voices pronounce Devanagari Hindi much better than Hindi written in Latin letters ("aapka" can come out oddly; "आपका" won't).

Speaking: pick a voice that sounds local

Choose a voice made for Indian Hindi and English rather than an American voice with a Hindi mode. Listen for three things before you decide:

1. Does it say Indian names correctly?
2. Do rupee amounts and dates sound natural ("ek hazaar chaar sau ninyanve rupaye")?
3. Does the switch between Hindi and English words sound like one person?

Numbers, dates and money

Callers notice numbers more than anything else. Ask the agent to say amounts and times the way people say them, and to read phone numbers and pincodes in small groups. If your tools return raw values ("2026-10-14T11:30"), tell the agent to say them naturally ("Tuesday, saadhe gyarah baje").

Test with real speech

Text tests won't catch code-mixing problems. Call the agent in the browser and speak the way your customers do, including mid-sentence switches, background noise and short answers like "haan" and "theek hai". Read the transcripts afterwards: if the transcriber is spelling Hindi words badly, add keyterms or switch the transcriber before you change the prompt.