Skip to main content
Early accessGnani Evon v3.5 is now available as an API to early access registrants.
A voice agent runs three models in a loop: Prisma turns the caller’s speech into text, Evon writes the reply, and Timbre speaks it. This recipe wires them together in LiveKit or Pipecat.

When to use this

  • You want a phone or web voice agent that speaks Indian languages.
  • You already use LiveKit Agents or Pipecat, or want a framework to handle audio, turn-taking and interruptions.
  • You want the reply to start playing before the whole answer is written.

Check the LLM settings first

Before you build the agent, send one request with the settings a voice agent needs: thinking off, streaming on and short replies.
The reply streams in a few pieces. Joined, it reads something like:

Build the agent

The Gnani LiveKit and Pipecat plugins provide speech-to-text and text-to-speech only. For the LLM, use each framework’s OpenAI plugin with the Evon base URL. This works because the API is OpenAI-compatible.
Install the packages:
Set GNANI_API_KEY, plus LIVEKIT_URL, LIVEKIT_API_KEY and LIVEKIT_API_SECRET for your LiveKit project. Save this as agent.py:
agent.py
Download the voice activity model once, then talk to the agent in your terminal:
Use python agent.py dev to connect it to your LiveKit project instead.

How it works

  1. The framework’s voice activity detection decides when the caller has finished a turn.
  2. Prisma’s final transcript is added to the conversation as a user message.
  3. The framework sends the whole conversation to Evon with stream: true.
  4. As text arrives, the framework sends it to Timbre a sentence or phrase at a time, so audio starts before the reply is finished.
  5. If the caller interrupts, the framework stops the audio and cancels the reply.

Options


Gotchas

One request at a time on Free and Pay-as-you-go plans. A voice agent that starts a new reply while the last one is still streaming, for example after an interruption, can get 429 RATE_LIMITED. Make sure the old reply is cancelled first, and see Limits for your plan.
  • Thinking can empty the reply. If thinking is left on, it can use the whole max_tokens budget and the reply comes back empty with finish_reason: "length".
  • Every turn resends the conversation. Long calls use your daily token quota faster. Trim old turns in long calls.
  • Personal data in replies is masked. If the model would say a phone number, the text contains [REDACTED_PHONE], and the TTS would read that out. Have your own code speak such details. See Safety and privacy.
  • Say it is an AI. Tell callers at the start that they are talking to an AI assistant, as the greeting above does.
Any framework that can call an OpenAI-compatible Chat Completions API can use Evon. Set the base URL to https://api.vachana.ai/v1, the model to evon-v3.5 and the API key to your Gnani key, and send reasoning_effort: "none" for voice.

LiveKit plugin

Prisma and Timbre settings for LiveKit Agents.

Pipecat plugin

Prisma and Timbre services for Pipecat pipelines.

Streaming

How streamed replies arrive, chunk by chunk.

Thinking

When to turn thinking off, and what it costs.