Full-duplex S2S
Listen and speak on the same timeline. Interruptions and overlap included.
Samvad‑v1
Full-duplex speech for Indic languages. One model to talk, listen, think, and act.
Model
One end-to-end speech model, not a cascade. Full-duplex bidirectional audio: no push-to-talk, no awkward half-duplex pauses while it thinks.
While you speak, it can reason, plan, and invoke tools in the background, then respond in voice without dropping the live stream. Built for Indic languages, accents, and code-switching.
Request early accessParallel lanes
Listen
Always on
Talk
Same time
Think + tools
Background actions
Fewer moving parts than a cascade. No half-duplex waiting.
Why it matters
Built for production voice agents that need to feel human, not turn-based.
Listen and speak on the same timeline. Interruptions and overlap included.
Skip ASR→LLM→TTS glue. One model, fewer failure modes, simpler ops.
Reason over the live stream instead of freezing until a turn ends.
Look up, book, fetch. Tool calls run without killing the conversation.
Accents, code-switching, and regional flow. Not English with a Hindi skin.
WebSocket API and SDKs. Integrate duplex voice without three services.
Languages
Eight languages at v1: Hindi, Tamil, Telugu, Kannada, Malayalam, Bengali, Marathi, and Sanskrit, with more in training.
Sanskrit
Native Sanskrit S2S for liturgy, academia, and preservation at launch.
About
We research full-duplex speech-to-speech for Indic languages, the kind of live dialogue cascade pipelines and half-duplex bots still cannot deliver.
Samvad (संवाद): dialogue and resonance. That is the bar for every model we ship.
संवाद
dialogue · conversation · resonance
Early access for teams shipping real-time Indic speech, not another cascade.