Skip to main content
GoogleLiveProvider adapts Gemini Live to VoiceAgent. It uses the optional Google SDK and resolves GOOGLE_API_KEY when connecting.
The v4 adapter default is gemini-3.8-live. This is a native realtime model, separate from the google(...) text model provider. Your account must have access to the configured Live model.

Media and tools

Connect the returned VoiceAgent using an authenticated identity, subscribe to events, then supply mono PCM input at 16 kHz. Output is PCM at 24 kHz. The host captures, plays, and converts audio; the adapter does not open your microphone or speaker. Gemini owns tool continuation at the provider connection. Let VoiceAgent coordinate local tool execution rather than issuing competing response commands. Provider-specific options are validated before opening a session.

Reconnection

Recovery is opt-in. Gemini can resume only from a usable safe provider checkpoint. A resumption handle does not authorize replay of new audio or tool results. Unsafe checkpoints or uncertain tool effects stop recovery; see voice recovery. Close the session when the call ends. If your application needs separate recognition/synthesis providers and coordinator-owned history, choose StreamingVoicePipeline instead.