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

# Realtime

> Speech-to-speech providers that replace the voice and transcriber

A realtime provider handles speech in and speech out with a single model. Instead of transcribing the caller, sending text to a language model, and speaking the reply, one model does all three at once.

The result is a faster reply, and a model that hears tone and pacing rather than just words.

## Supported providers

| Provider                                                               | Notes                                                |
| ---------------------------------------------------------------------- | ---------------------------------------------------- |
| [OpenAI Realtime](/talk/configurations/providers/openai-realtime)      | Uses an ordinary OpenAI key                          |
| [Gemini Live](/talk/configurations/providers/google-gemini-live)       | Through Google AI Studio                             |
| [Vertex AI Gemini Live](/talk/configurations/providers/vertex-ai-live) | The same model through your own Google Cloud project |
| [Grok Realtime](/talk/configurations/providers/grok-realtime)          |                                                      |
| [Ultravox](/talk/configurations/providers/ultravox)                    |                                                      |

## What changes when you use one

<Warning>
  A realtime provider replaces the voice and transcriber, but you **must still configure an LLM**. It runs the work the realtime model does not: extracting variables and running QA analysis after the call. Without one, those features silently produce nothing.
</Warning>

You also give up mixing providers. With a standard setup you can pair one company's language model with another's voice; a realtime model brings its own voice and its own behaviour, and you take them together.

Context compaction does not apply in realtime mode and is ignored if enabled.

## Turning it on

Enable **Realtime** in [Model Configurations](/talk/configurations/inference-providers) to use it everywhere, or switch a single workflow to realtime from its Model Overrides tab.

Credentials come from a [connection](/talk/configurations/connections), same as any other provider.