Most Indian callers don't speak textbook Hindi — they mix English product names, numbers, and phrases into Hindi sentences mid-thought. Sarvam's speech stack is trained on exactly that: natural Hindi-English code-switching, transcribed as it was spoken instead of being forced into one language or the other.
The result is an agent that hears "mera EMI Tuesday ko due hai" as one utterance, not a broken Hindi fragment followed by English noise — and answers in the same register the caller used.
Caller
"Mera loan statement email kar do, please."
Pipeline
Sarvam transcribes the mixed utterance as-is — no forced Hindi-or-English split.
Agent
Replies in the same code-mixed register and completes the request.
Built on Sarvam's speech stack — tuned for how India actually speaks, including code-mixed Hinglish and Indian-accented English.
Full STT + TTS with natural prosody
Code-mixed Hindi-English, natively
Recognition and synthesis
End-to-end Tamil calls
Natural synthesized voices
Maharashtra-first deployments
Commerce and support flows
Karnataka-facing call flows
Kerala-facing service lines
Indian-accented local voices
Odisha callers in their language
Tuned to Indian names and numbers
The same agent platform covers 40+ global languages through your configured STT and TTS providers — including:
Real conversations don't stay in one language. A caller might start in Hindi, drop into English for a product name, then switch back — and the dedicated language-switch processor in the pipeline tracks that automatically.
The agent follows the caller instead of forcing an upfront language menu. No "press 1 for Hindi", no restarting the call — the response simply continues in the language the caller is using now.
Caller speaks Hindi
Agent detects Hindi and responds in Hindi.
Caller switches to English
The language_switch_processor detects the change mid-conversation.
Agent follows automatically
Responses switch languages without the caller repeating themselves or pressing a key.
STT, TTS, and telephony are configurable per agent — Sarvam for Indian languages, Deepgram, ElevenLabs, and Cartesia for global ones.
| Layer | Indian languages | Global languages |
|---|---|---|
| STT (speech-to-text) | Sarvam saaras:v3 | Deepgram nova-3, Azure Speech, Google STT |
| TTS (text-to-speech) | Sarvam bulbul:v2 | ElevenLabs, Cartesia |
| Telephony | 8 providers with failover | Exotel, Vobiz, Twilio, Plivo, Telnyx, Vonage, Cloudonix, Asterisk ARI |
Provider choice affects turn latency — see the stage-by-stage breakdown on Benchmarks.