Skip to content

GPT-Live / OpenAI Realtime ​

qwen-audio-agent can connect to OpenAI's GPT-Live / Realtime WebSocket API as a cloud voice frontend. Gateway still owns frontend tools, memory, reminders, task delegation, and backend-Agent orchestration; the provider adapter only translates the OpenAI GA Realtime wire shape into the shared runtime.

Configuration ​

Edit the user configuration file shown by qwenaudio config:

dotenv
QWEN_AUDIO_REALTIME_PROVIDER=gpt-live
OPENAI_API_KEY=your-openai-key
Optional settingDefaultDescription
GPT_LIVE_REALTIME_URL / OPENAI_REALTIME_URLwss://api.openai.com/v1/realtimeWebSocket endpoint
GPT_LIVE_REALTIME_MODEL / OPENAI_REALTIME_MODELgpt-realtime-2.1Current built-in model profile
GPT_LIVE_REALTIME_VOICE / OPENAI_REALTIME_VOICEEmptyService default, or a voice ID supported by the model
GPT_LIVE_API_KEYEmptyAlias for OPENAI_API_KEY when the realtime frontend needs a separate credential

Desktop exposes the same fields under Voice frontend → GPT-Live. Restart a terminal Gateway after editing the file; for an installed service, run qwenaudio gateway restart.

Integration boundary ​

  • The adapter uses OpenAI's GA Realtime dialect over WebSocket. It sends the model as the URL model query parameter and authenticates with Authorization: Bearer ....
  • Input and output use mono 24 kHz PCM. Client audio is resampled by the Gateway client before it reaches the provider.
  • Gateway registers its own function tools with the model. Built-in OpenAI tools are not enabled by this adapter.
  • Response metadata correlation is enabled, so Gateway-created replies can be distinguished from automatic server-side responses.
  • The current built-in model profile supports text/audio input and text/audio output. Live visual frames are not negotiated for this provider.

Validation boundary ​

Local protocol tests cover URL construction, session configuration, model capability reporting, and provider registration. Live model behavior, voices, latency, quotas, and regional availability require a valid OpenAI account and are governed by OpenAI's current Realtime API documentation.