OpenAI Realtime

An OpenAI voice model hears and speaks directly, while your configured agent handles files, tools, private context, current information, and longer work. You can keep talking while the specialist works.

GPT Realtime and GPT Live are separate

Both appear under OpenAI Realtime → Model in settings, but Sleeptalk uses different implementations for them:

For continuous browser calls, read GPT Live. For several concurrent Sleeptalk workers, use Cascade or an eligible GPT Live browser profile; see background workers. The shared provider category does not make these routes interchangeable.

Set up GPT Realtime

1. Open sleeptalk settings and choose OpenAI Realtime as the speech service.
2. Enter an OpenAI API key privately through the settings page. The key needs billing and access to the selected model.
3. Choose a GPT Realtime model, your voice, and your language. Retain your configured agent for workspace and tool work.
4. Save and start a new call. Use a separate profile if you want to compare it with Cascade or GPT Live.

In profiles this is route.type = "hybrid", with an OpenAI model in [route] instead of separate [stt] and [tts] tables. The hybrid label is an internal route name, not a fourth speech product. See the repository guide for a scripted setup example.

What to expect

The voice model can answer conversational and general-knowledge questions. Requests that need your workspace, tools, the web, or private context go to your agent, and the voice reports what it found. Realtime can begin a reply before the final transcript arrives; this is different from a Cascade provider that uploads a completed utterance.

There is one specialist task at a time. A new specialist request cancels the previous one; ordinary conversation can continue while it runs. Interrupting spoken output and cancelling work have separate lifetimes. Check the task's state before assuming its tools stopped, and remember that cancellation does not undo completed actions.

OpenAI voice usage and your agent's model usage are separate charges. Check the selected model's current provider pricing; timings and costs from an old research trial do not describe every current call.

Support and verification

The repository records live GPT Realtime checks in macOS terminal calls and iPhone-to-Ubuntu-VPS browser calls over tailnet and hosted links on September 25, 2026. Discord remains preview. This replaces this page's earlier research-only description; it does not extend that validation to GPT Live.

Use the support matrix for the release contract. Real microphone, interruption, and spoken-result checks remain human checks.