> ## Documentation Index
> Fetch the complete documentation index at: https://docs.praxis-ai.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Gemini Live Voice

> Enable real-time voice conversations powered by Google Gemini's native audio WebSocket API.

Gemini Live brings real-time, natural voice conversations to your Digital Twin using Google's native audio API. Unlike text-based approaches that convert speech to text and back, Gemini Live processes audio natively — resulting in lower latency and more natural-sounding responses.

<Tip>Gemini Live is ideal for interactive learning, tutoring, and conversational scenarios where low-latency, natural dialogue is important.</Tip>

***

## How It Works

<CardGroup cols={3}>
  <Card title="User Speaks" icon="microphone">
    Audio is captured from the user's microphone and streamed over a WebSocket connection directly to Google's Gemini API.
  </Card>

  <Card title="Gemini Processes" icon="brain">
    Gemini processes the audio natively — understanding speech, generating a response, and synthesizing audio output in a single model pass.
  </Card>

  <Card title="User Hears" icon="headphones">
    The AI's audio response streams back in real time. Both sides of the conversation are transcribed for the chat history.
  </Card>
</CardGroup>

```
User speaks → WebSocket → Gemini (native audio processing) → Audio response → User hears
                                    ↕
                            Tool calls (search, files, web)
```

***

## Prerequisites

Before enabling Gemini Live, ensure:

* A **Gemini API key** is configured (at the instance or environment level)
* The instance has a **Gemini-compatible model** assigned as the conversation model
* The **Convo mode voice provider** is set to Gemini Live in the instance configuration

**Supported Gemini Live models:**

| Model                                           | Status      |
| ----------------------------------------------- | ----------- |
| `gemini-3.1-flash-live-preview`                 | Default     |
| `gemini-2.5-flash-native-audio-preview-12-2025` | Alternative |

***

## Admin Setup

<Steps>
  <Step title="Configure a Gemini API Key">
    Gemini Live uses a 2-tier API key resolution:

    1. **Instance level** — The `gemini_api_key` field in the **Configuration** tab (highest priority)
    2. **Environment level** — The `GEMINI_API_KEY` environment variable (lowest priority)

    At least one of these must be set for Gemini Live to function.

    <Note>This differs from standard Gemini chat, which also checks the AI model's `api_key` field. Gemini Live does not use per-model API keys.</Note>
  </Step>

  <Step title="Set the Voice Provider">
    In the **Personalization** tab under the **Convo** section, set the **Convo mode voice provider** to **Gemini Live**. This tells Pria to use the Gemini WebSocket API instead of OpenAI Realtime or ElevenLabs.
  </Step>

  <Step title="Choose a Default Voice">
    Select a default voice for your instance from the `gemini_rt_voice` setting. Users can override this in the Convo Mode panel.

    **Main voices:**

    | Voice  | Style       |
    | ------ | ----------- |
    | Puck   | Upbeat      |
    | Charon | Informative |
    | Kore   | Firm        |
    | Fenrir | Excitable   |
    | Aoede  | Breezy      |
    | Zephyr | Bright      |
    | Leda   | Youthful    |
    | Orus   | Firm        |

    <Info>Additional voices are available in the voice selector dropdown when a user opens Convo Mode. [Preview Gemini voices on Google AI Studio](https://aistudio.google.com).</Info>
  </Step>
</Steps>

***

## Features

<AccordionGroup>
  <Accordion title="Tool Calling" icon="wrench">
    During a voice conversation, the AI can invoke tools just like in text mode — searching the web, looking up files in IP Vault, querying Canvas, and more. Tool results are incorporated into the audio response naturally.
  </Accordion>

  <Accordion title="Live Transcription" icon="closed-captioning">
    Both the user's speech (input) and the AI's spoken responses (output) are transcribed in real time and displayed in the Convo Mode panel. Transcripts are saved to conversation history for later review.
  </Accordion>

  <Accordion title="Proactive Audio" icon="comment-dots">
    When enabled, the AI can initiate speech without waiting for the user to speak first — for example, offering follow-up suggestions or asking clarifying questions. This creates more natural turn-taking in tutoring scenarios.
  </Accordion>

  <Accordion title="Noise Reduction" icon="volume-low">
    A noise reduction toggle filters background audio from the user's microphone, improving recognition accuracy in noisy environments like classrooms or open offices.
  </Accordion>

  <Accordion title="Text Input During Voice" icon="keyboard">
    Users can type messages during an active voice conversation. The typed text is sent alongside the audio stream, and the AI responds with both voice and text.
  </Accordion>

  <Accordion title="Ephemeral Token Security" icon="shield-halved">
    Each voice session uses a short-lived ephemeral token generated by the server. No long-lived API keys are exposed to the browser. Tokens are created on-demand and expire after the session ends.
  </Accordion>
</AccordionGroup>

***

## User Experience

When a user opens Convo Mode with Gemini Live configured, they see:

1. **Voice selector** — Dropdown with available Gemini voices
2. **Start/Stop button** — Begins and ends the voice session
3. **Noise reduction toggle** — Filters background noise from the microphone
4. **Proactive audio toggle** — Allows the AI to speak without being prompted
5. **Text input toggle** — Enables typing during a voice conversation
6. **Live transcript** — Real-time display of both user speech and AI responses

***

## Troubleshooting

<AccordionGroup>
  <Accordion title="Voice session fails to start">
    * Verify a Gemini API key is configured at the instance or environment level
    * Check that the instance's voice provider is set to Gemini Live
    * Ensure the browser has microphone permissions granted
  </Accordion>

  <Accordion title="Audio quality issues">
    * Enable the noise reduction toggle in the Convo Mode panel
    * Use headphones to prevent echo/feedback
    * Check your internet connection — Gemini Live requires a stable WebSocket connection
  </Accordion>

  <Accordion title="Tools not working during voice">
    * Confirm tools are enabled on the instance
    * Some tools may not be available during voice mode — check the admin Tools panel
  </Accordion>
</AccordionGroup>

***

## Related

* [Convo Mode User Guide](/mdx/user-guide/interface/convo-speech) — End-user guide for voice conversations
* [AI Models](/mdx/admin-guide/ai-models) — Configure Gemini models
* [ElevenLabs Integration](/mdx/integrations/elevenlabs/introduction) — Alternative voice provider
* [Configuration](/mdx/admin-guide/configuration) — Instance voice settings