AI calling in Hindi and other Indian languages

A dental clinic in Pune puts an AI agent on its phone line. The demo in English is flawless. Then a real caller says, "Mujhe Saturday ko evening mein appointment chahiye, root canal ke liye." The agent answers in stiff English, mishears "root canal", and asks the caller to repeat the date twice.

Nothing was wrong with the agent's English. The problem is that Indian phone callers rarely stay inside one language, and most voice AI is first tested on people who do. This post covers what to check before you go live.

Callers mix languages, and that is normal

Hinglish is the best-known case: Hindi grammar with English nouns, or the reverse. The same happens in Tamil and English, Telugu and English, Marathi and Hindi, Bengali and English. A caller may open in English because they expect a business to prefer it, then drift into their own language once the conversation feels easy.

For an agent this has two consequences. First, speech recognition has to cope with a sentence that changes language halfway through. Second, the agent has to decide which language to reply in, and that choice is visible to the caller within a second.

The agent here follows the caller's lead without comment. It does not announce a language change or ask which language the caller prefers. That is usually the right default.

Choosing recognition and voices for each language

Speech recognition is the stage most likely to fail you. Vendors differ widely in how well they handle Indian accents, code-switching and phone-line audio, which is compressed and narrow compared with a studio microphone. Published benchmarks rarely match your callers, so treat a vendor's headline accuracy as a starting point only.

Three practical points:

• Pick recognition per language, not per platform. A vendor strong in Hindi may be weaker in Malayalam or Odia. Platforms that let you choose a vendor for each stage make this possible; a single all-in-one model gives you less control.
• Listen to the voice in the caller's own dialect. A Marathi voice that sounds natural in Pune may sound foreign in Nagpur. Play samples to two or three staff members from the region you serve.
• Check that the voice handles English words inside Indian sentences. Some voices read "appointment" in the middle of a Hindi sentence with an accent that jars. Others handle the switch cleanly.

Numbers, dates, addresses and names break first

Ordinary conversation is the easy part. The failures cluster around structured information.

• Numbers. Callers say phone numbers digit by digit, in pairs ("nau-aath"), or with "double" and "triple". Amounts are said in lakhs and crores. Ask the agent to repeat back any number it will act on, in the grouping the caller used.
• Dates. "Parson", "agle Monday" and "is mahine ki 15 tareekh" all need to resolve to a real calendar date. The agent should confirm the full date aloud ("Saturday, 31 October") before booking.
• Addresses. Landmarks matter more than street numbers: "Gokhale Road, Sai Mandir ke saamne". Do not expect clean structured data. Capture what the caller says, and have a person verify before dispatch.
• Names. A recogniser trained on common words will mangle surnames and spell them in whichever script it leans towards. For anything that must be exact, ask the caller to spell it or send the spelling by WhatsApp afterwards.

Scripts in transcripts

A transcript of a Hindi call can come back in Devanagari, in Roman letters, or as a mix, depending on the recogniser and its settings. This matters more than it seems. If your CRM search, your keyword alerts or your staff expect Roman script and the transcript is in Devanagari, nobody finds anything.

Decide one convention early, such as storing the original script and also a Roman version for search. Then read twenty real transcripts to see whether English words inside Hindi sentences are written in Roman or transliterated into Devanagari, because that varies and it changes how you write rules over the text.

When the agent should switch language

Three rules cover most cases.

1. Reply in the language the caller used in their last full sentence, unless they ask otherwise.
2. Do not switch on a single borrowed word. A caller who says "appointment" in a Hindi sentence is still speaking Hindi.
3. If the caller asks, "Can you speak Tamil?" or "English mein bolo", switch at once and stay switched.

Cora Call agents speak 15 languages, including Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia, and can switch language mid-call. Whatever platform you use, write the switching rule into the agent's instructions rather than assuming it will behave sensibly.

Politeness, register and what callers expect

Hindi and several other Indian languages carry levels of formality in the verb itself. An agent that says "tum" to a customer in their sixties sounds rude; one that says "aap" to everyone is safe. Tamil, Marathi and Bengali have similar distinctions. Put the register you want in the prompt, and check the voice keeps it.

Greetings matter as well. Many callers expect "Namaste" or "Vanakkam" and a business name before anything else, and a flat "How can I help?" in the wrong language sets the tone for the whole call. Read your own greeting aloud to someone from the region and ask whether it sounds like a local business or a foreign one.

Test with real callers before launch

Staff reading from a script are the worst test group, because they speak slowly and clearly. Use people who sound like your customers.

Test in noisy places too: a market, a moving auto-rickshaw, a speakerphone in a shop. Phone audio from these settings is what your agent will actually hear. Our post on latency in AI phone calls explains why a slow reply makes a language slip feel worse, and you can hear an agent for yourself on the live demo.

Where to start

• List the two or three languages that cover most of your callers, from your own call logs.
• Pick recognition and a voice for each, and have regional staff listen to the voice.
• Write the language-switching rule and the confirm-back rule for numbers and dates into the prompt.
• Decide how transcripts are stored and searched.
• Run a pilot with real callers on a single line, and read the transcripts weekly for the first month.