For developers
Your voice application. Your conversation rules.
Build a voice experience around your users: open conversation when they need freedom, guided steps when the task needs structure, and configurable reflex answers.
Human-like voice agents. 11 Indian languages and English.
Begin with a session.
The Swarit developer integration starts with a request from your backend. Supply instructions, a language, a voice, and a duration limit. Wait for the session to become ready, then request temporary connection credentials for your client.
The application can publish microphone audio and receive the agent’s audio through the returned LiveKit transport. Keep permanent API credentials on your backend. Temporary connection credentials are what you pass to the browser.
Choose how the conversation moves.
An open conversation can help someone explain a problem in their own words. A guided conversation can help collect the details a task requires. In the API preview, leave steps empty for an open conversation or supply a validated conversation graph.
Reflex answers are a separate option. They can be enabled or disabled independently of steps. With reflex answers disabled, the backend supplies the reply; with guided steps enabled, the backend still validates progression and captured information.
Make responsiveness something you test.
Speed matters because people feel the gap between speaking and hearing an answer. For your application, measure the complete experience: session startup, readiness, the first audible response, and responses throughout the conversation.
Try short questions, interruptions, corrections, and different network conditions. Record typical and slower responses. We are not publishing a fastest-model claim here: a useful latency comparison needs the same workload, serving conditions, and measurement method.
Three prototypes worth exploring.
Build a practice partner that lets learners repeat an interview answer and try again. Keep the conversation open so they can change direction.
Build an appointment intake assistant with steps for the requested service, preferred time, and a final confirmation. Let users correct an earlier answer.
Build a voice-first form where a person can explain several details in one sentence. Check whether the structured result matches what they actually said. These are starting points for your application, not prebuilt integrations.
Connect the experience to a clear cost.
Realtime developer usage is priced at ₹2 per connected minute. Choose a maximum session duration and measure how long real users need to complete the task. Track useful outcomes alongside minutes.
Read the documentation for the integration preview and try the public playground. Standalone TTS and STT are coming soon. Public paid realtime API access remains in preparation; the current docs explain the interface you can plan around.