"Bhai, kal ka appointment confirm karna tha, but I'm not sure what time works. Evening mein free ho?"
That's a completely ordinary sentence for a huge number of Indian callers, and it's also the sentence that breaks a lot of voice AI. Hindi and English share the sentence, switching mid-thought, sometimes mid-phrase, without the speaker noticing they did it at all.
If you've spent time on the phone with Indian customers, you already know this. Gujarati speakers mix in English nouns for prices, dates, and brand names the same way, and plenty of callers open in English and drop into their first language the moment the question gets specific. Two languages, one sentence, no signal that a switch is coming.
For a business, this has a direct cost. A clinic, salon, or coaching institute whose phone agent can't follow a mixed-language caller rarely gets a complaint; the agent just sounds slightly off, and callers hang up and try someone else.
If you're deciding whether a voice agent can handle your callers, the comparison table and the five-minute test further down will tell you most of what you need.
What is code-switching in voice AI?
Code-switching is the natural mixing of two languages in one conversation. In India the most common pairing is Hindi and English, often called Hinglish, but Gujarati-English and many other combinations follow the same pattern.
It shows up at every level:
- Between sentences: "Appointment kal hai. Can you confirm the time?"
- Within a sentence: "Saturday ko slot hai kya?"
- Inside a phrase: "colour ke liye kitna time lagega?"
For voice AI, handling it means three things working together: recognizing the words correctly across the switch, understanding the request as one thought, and replying in a language and register that fit the caller.
Aiva Chat supports 13 languages, including Hindi, Gujarati, and English, and can follow the caller or run in a two-language mode. That covers the most common Indian pairings our customers see, but not every Indian language, and the limits matter as much as the coverage.
How Aiva Chat follows a code-switched call
Step 1: The greeting sets the starting language
In two-language mode, the call opens in your chosen first language, typically the one most of your callers expect. In follow-the-caller mode, the assistant listens and matches whatever the caller speaks.
Step 2: The caller's speech is understood as one request
Aiva Chat's voice runs on OpenAI's speech-to-speech Realtime models. OpenAI's Realtime API documentation describes them as built for voice agents with low first-audio latency and natural turn-taking. A mixed sentence is processed as speech, not chopped into separate languages first.
Step 3: The assistant switches when the caller does
When a caller moves from Hindi into English, or settles into Hindi after an English greeting, the assistant follows. In two-language mode it moves to the second language when the caller does.
Step 4: It answers both halves of the question
A caller who asks "colour ke liye kitna time lagega, aur price kya hai?" gets the time and the price in one reply, from your own service information, rather than an answer to whichever half was recognized.
Step 5: Interruptions work the same way
Callers cut in to correct a date or change their mind. Interruption sensitivity applies to mixed-language calls exactly as it does to single-language ones.
Why handling code-switching well matters
Callers don't have to perform for the machine
When a voice agent can't follow mixed speech, callers start speaking slowly in "proper" Hindi or English, or give up. Following the caller's natural speech removes that friction, which is the whole point of answering by voice.
The tone fits the caller
A caller who opens with "bhai" is signaling a casual conversation. A stiff, textbook-formal reply sounds like a script even if every word is correct. Replying in the caller's mix and tone makes the assistant sound like it listened.
Urgent calls get the right tone
A patient describing symptoms in a rushed mix of Hindi and English needs a reply that matches that urgency. A serene, formal answer feels like the system processed words without hearing the person.
No language menu
"Press 1 for English, 2 for Hindi" forces a choice many bilingual callers can't cleanly make. We've written about why press 1 for English loses customers; following the caller avoids the menu altogether.
Consistent across channels
On the web widget, replies follow the visitor's language too, and when staff click Suggest a reply in the SMS inbox, the draft comes back in the customer's language for them to check and send. Customer-facing booking pages, confirmations, and reminder messages are available in Hindi and Gujarati as well as English, so the written side of a booking can match the language the call was in.
Code-switched calls in practice
Four composite calls, the kind Indian front desks take daily:
The appointment confirmation
Problem: A caller asks to confirm tomorrow's appointment, unsure about timing, wanting an evening slot, all in one Hinglish sentence. How Aiva Chat handles it: It answers about the evening slots that are actually free, in the same casual mix, rather than a stilted "Aapka appointment kal hai. Kripya samay bataen." Outcome: The caller books an evening slot without repeating themselves.
The clinic availability check
Problem: A caller asks "doctor available hai kya abhi?" How Aiva Chat handles it: It answers directly from the clinic's hours and open slots, without routing through a language menu. Outcome: A direct answer in the caller's own register.
The two-part salon question
Problem: A caller asks about colour treatment time and price in one mixed sentence. How Aiva Chat handles it: It gives both answers from the salon's price list in a single reply. Outcome: No half-answered question, no follow-up call.
The Gujarati-English business call
Problem: A wholesaler's customers open in Gujarati and switch to English for product names and quantities. How Aiva Chat handles it: In two-language mode with Gujarati first and English second, the assistant follows the switch. Outcome: Orders and enquiries are captured accurately in the caller's natural speech.
How approaches to multilingual voice compare
| IVR language menu | Single-language model | Translation layer | Follow-the-caller voice (Aiva Chat) | |
|---|---|---|---|---|
| Handles mixed sentences | No | Poorly | Poorly, guesses one language | Yes |
| Caller must choose a language | Yes | Effectively, yes | No | No |
| Matches the caller's register | No | Rarely | No, outputs "clean" language | Aims to |
| Adds delay | Menu time | No | Extra processing step | No separate path |
| Coverage of every Indian language | Depends on recordings | One language | Often broad | 13 languages, Hindi and Gujarati among them |
An IVR menu is predictable and can cover any language you record prompts for, which is its one real advantage. But it forces a choice, and it can't understand anything beyond keypad presses.
A single-language model sounds natural in its own language and stumbles at the seams. A translation layer offers broad nominal coverage but resolves everything into one language, losing the meaning carried by the mix. The W3C's Internationalization Activity publishes guidance for handling multiple languages on the web because these problems are real even in text, before speech adds its own difficulties.
Following the caller handles the common Indian pairings well, within the 13 languages Aiva Chat supports (see the pitfalls below for what that leaves out). Our piece on regional-language AI vs. translation goes deeper into why translation struggles.
How to set up language handling in Aiva Chat
Step 1: Listen to how your callers talk
Before changing any settings, listen to a handful of real calls or think about your front desk's last week. Which languages do callers open in, and which do they switch into?
Step 2: Pick the language mode
On the Voice & Speech page, choose "Follow the caller," "Always one language," or "Two languages." For a Hindi-English customer base, two-language mode with the more common opening language first is usually right.
Step 3: Write the greeting in the first language
Write the greeting in the language most callers expect, and make sure it says they're speaking with an AI assistant. Keep it short so callers can start talking quickly.
Step 4: Tune turn-taking
Callers choosing words in a second language often pause mid-thought. If the assistant cuts in too early, adjust the listening and turn-taking settings or interruption sensitivity.
Step 5: Test in the real mix
Call it as your customers would: open in English, switch to Hindi for a date, interrupt once, ask a two-part question. Repeat in Gujarati-English if relevant.
Step 6: Review transcripts
Read calls from the first weeks, especially mixed-language ones, and add corrections where an answer missed part of a question.
Common pitfalls with mixed-language voice AI
Pinning one language for a bilingual audience
"Always Hindi" sounds safe until half your callers open in English. For mixed customer bases, follow the caller or use two-language mode.
Testing only in clean language
A demo in perfect English or perfect Hindi proves little. Your callers will mix, so your test calls should too.
Writing knowledge only in formal language
If your service names and prices are written in formal Hindi only, answers may come out formal. Write entries the way your staff actually explain things, including the English terms customers use.
Assuming every Indian language is covered
Aiva Chat supports 13 languages, not every Indian language. If your customers speak Tamil or Bengali, Aiva Chat isn't the right fit for those calls today.
Ignoring accents alongside language
Language mixing and accent are separate challenges. Our post on whether AI can understand Indian accents covers the accent side.
Best practices for code-switching voice AI
- Choose the mode from real call patterns, not assumptions.
- Open in the language most callers expect.
- Write knowledge in everyday language, including common English terms.
- Test with mixed sentences and interruptions, not clean scripts.
- Tune turn-taking for callers who pause to find words.
- Be honest with customers about languages you don't cover.
- Review mixed-language transcripts during the first month.
Conclusion
For a business with Hindi-English or Gujarati-English callers, language mode is the most consequential voice setting you'll pick, more than the voice itself. Two-language mode with the usual opening language first suits most Indian customer bases; follow-the-caller suits a wider mix; pinning one language rarely sounds right to bilingual callers.
The decision point is coverage. If your callers mostly mix Hindi or Gujarati with English, Aiva Chat fits. If a meaningful share speak Tamil, Telugu, or Bengali, those calls won't get the same experience, and you should weigh that before switching your main line. The languages page has the full list.
Either way, judge it by ear: call it in your customers' mix and listen for whether it keeps the thread.
Ready to hear it follow your callers?
Set two-language mode, write a one-line greeting, add your services and prices, then call it the way your regulars do, "kal ka slot" and all.
The language modes are on every plan, from Starter at $49 a month; your OpenAI and Twilio usage is billed by those providers on your own accounts. If you'd like to hear it before setting anything up, call +1 (970) 972-0974.
Start your 14-day free trial with no card required, or book a demo to try mixed-language calls with our team.
Posts from the people building Aiva Chat in Rajkot — how the product works, what we have learned building it, and guides for getting the most out of an AI receptionist.