Hinglish AI voice agents: handling code-switching on real calls

Updated 2026-09-19 · 5 min read · By the Convo360 team

Why Hinglish is harder than 'just support Hindi'

Most Indian callers don't speak pure Hindi or pure English, they switch mid-sentence: 'mera order kab tak deliver hoga, address change kar sakte hain kya?' A voice bot that only recognises one language at a time will mishear the switch, or worse, respond in the wrong language entirely and sound like it wasn't really listening.

This is different from simply 'adding a Hindi language option' in a settings menu. Code-switching support means the speech-to-text and the AI reply both need to handle a single utterance containing both languages naturally, the way a bilingual call-centre agent does without even noticing they're doing it.

This shows up constantly in real business calls. A customer confirming a delivery address in English while asking about payment ('cash on delivery hoga kya?') in Hindi within the same breath, or a caller starting formally in Hindi and slipping into English for a specific product name. A bot built only around clean, single-language sentences will stumble on exactly this pattern.

Where accuracy actually breaks down

• Speech-to-text (STT): the biggest failure point. Weak STT engines either force-transcribe Hindi words as English-sounding nonsense, or drop the English words entirely, garbling the transcript before the AI even sees it.
• Accents and regional pronunciation: a Mumbai caller and a Lucknow caller say the same Hinglish sentence differently, and a rural or older caller may speak more Hindi-heavy than a younger urban customer, so the STT needs to be robust to all of these, not tuned to one accent.
• The LLM reply itself: even with perfect transcription, a model with a rigid prompt ('always reply in English') will sound stiff or ignore the Hindi half of the question, producing an answer that's technically correct but socially odd.
• Background noise on the caller's side: a call from a shop counter or a moving auto-rickshaw adds another layer of difficulty on top of the language mixing itself.

Getting the voice bot right means fixing all of these layers together, not just picking a 'Hindi voice' from a dropdown and assuming the rest follows.

Prompt rules that make replies sound natural

The system prompt matters as much as the model. A few rules that consistently help:

• Explicitly instruct the agent to reply in the same mix of Hindi and English the caller used, rather than defaulting to one language for the whole call.
• Give a few example exchanges in the prompt itself showing natural Hinglish phrasing, not textbook Hindi, since the model mirrors the style of the examples you give it.
• Avoid instructing the model to 'translate' the caller's words, since that produces stilted, overly formal replies. Instead instruct it to respond naturally in the caller's own register and vocabulary.
• Keep sentences short. Long, grammatically complex Hinglish sentences are harder for both STT and TTS to handle cleanly than shorter, conversational ones.

Convo360 agents can also auto-detect the caller's language and switch mid-call, so if a caller starts in English and drifts into Hindi halfway through, a very common pattern once a caller relaxes into the conversation, the agent follows rather than forcing a single language for the entire call.

Choosing a voice for Hinglish

Not every text-to-speech voice handles code-switched text well. Some read English words in the sentence with an odd, over-anglicised accent, which sounds jarring to an Indian ear and breaks the illusion of a natural conversation. India-native voice providers like Sarvam and Smallest.ai are generally built with this mixing in mind, while OpenAI and ElevenLabs voices work well for English-heavy scripts and international audiences.

Test a sample Hinglish paragraph, ideally one pulled from a real customer message rather than something you wrote yourself, in the Voice Library before committing to one voice for a live agent. If none of the stock voices sound quite right for your brand, consider voice cloning from 1-2 minutes of audio to get a specific, familiar voice that your regular callers might already recognise from your storefront or call centre.

Setting this up and testing it on Convo360

1. Create a voice agent and pick a Sarvam or Smallest.ai voice suited to Hinglish delivery.
2. In the system prompt, add an explicit instruction to mirror the caller's Hindi-English mix rather than translating to one language.
3. Enable language auto-detection so the agent can switch mid-call if the caller does.
4. In Call behaviour, set Acknowledgement style to a natural short word like 'haan' or 'sure' rather than a stiff formal phrase, since small details like this affect how natural Hinglish sounds far more than most people expect.
5. Call the number yourself using real, unscripted Hinglish sentences, not typed test messages, since spoken code-switching behaves differently from written text.
6. Review call transcripts in Flow Logs for any point where the STT clearly mis-heard a switch, and adjust the prompt, voice or acknowledgement settings accordingly before opening the line to real customers.

Common mistakes to avoid

A frequent mistake is writing the system prompt only in English and assuming the model will figure out Hinglish replies on its own. It often can, but the phrasing tends to sound translated rather than natural, so it's worth explicitly showing the model a couple of real Hinglish exchanges in the prompt itself.

Another common issue is testing only with typed messages during flow setup and never actually calling the number. Typed Hinglish and spoken Hinglish are not the same thing. Spoken language has pauses, filler words like 'matlab' and 'toh', and pronunciation quirks that a text-only test will never surface. Always finish testing with an actual phone call before handing the number to customers.

Finally, avoid picking a voice purely because it sounds good reading pure English marketing copy in a demo. Insist on hearing it read an actual Hinglish sentence with a product name or an address in it before deciding.

Step by step

1. Pick an India-native voiceIn the Voice Library, choose a Sarvam or Smallest.ai voice and preview it against a Hinglish sample sentence.
2. Write a mixed-language prompt ruleInstruct the agent to reply in the same Hindi-English mix the caller uses, with a couple of example exchanges.
3. Turn on auto language detectionLet the agent detect and follow the caller's language shift mid-call instead of locking to one language.
4. Test with real spoken HinglishCall the number yourself with natural, unscripted mixed sentences rather than reading from a script.
5. Review Flow Logs for mis-heard switchesCheck transcripts for STT errors at the Hindi-English boundary and refine the prompt or voice choice.

Example conversation

Frequently asked questions

Related reading

Available across India — see WhatsApp & AI calling automation by state and city.