Why Most Voice AI Fails at Hinglish, and What "Code-Mixed by Default" Actually Means

More than 250 million people in India are estimated to code-switch regularly — blending Hindi and English inside a single sentence, not alternating between two clean conversations. That’s not an edge case for a voice agent built for India. It’s the default input. Most voice AI products aren’t built for it, and the reason is architectural, not a matter of adding more training data.

Most speech-to-text systems run utterance-level language identification — the model decides which language it’s hearing once, for the whole audio segment, before it starts transcribing. On clean monolingual speech that’s fine. On a sentence that opens in Hindi and closes in English — or the reverse, mid-word — the model has already committed to the wrong language for half the sentence. What comes out is substitution and deletion errors on whichever language it guessed wrong, not a slightly-worse transcript.

The fix isn’t “more Hindi data” or “more English data.” A model trained on a large Hindi corpus and a large English corpus separately still doesn’t know the actual mixing patterns — which words get borrowed, where the switch tends to happen in a sentence, how an English word gets pronounced once it’s sitting inside Hindi phonology. None of that shows up in either monolingual dataset on its own. It only shows up in real code-mixed speech.

“Mera order abhi tak deliver nahi hua, can you check kab tak aayega” is an ordinary sentence for a huge share of Indian callers — not a stress test. A voice agent that silently drops to English-only or Hindi-only the moment it detects the “wrong” language mid-sentence doesn’t fail loudly. It just mishears half of what was said, which is worse: the call sounds like it went fine and the data underneath it is wrong.

Code-mixed speech — Hinglish, Tanglish, and the equivalent registers in the other five languages — is the register the agent is designed for by default, not a mode you switch into. That’s the reasoning behind an Indic-native stack: Sarvam’s STT/TTS is built against real Indian call data rather than adapted from an English-first model, and it’s one of the providers Decibyl runs on, alongside OpenAI, Google, and ElevenLabs. See how it works for the full setup, or hear it for yourself on the language samples on the homepage.