Gabay sa Gemini 3.1 Flash Live API: real-time na boses, camera, at screen share
Gumawa ng mababang-latency na multimodal agents gamit ang Gemini 3.1 Flash Live — WebSocket, native audio, camera/screen 1 FPS, ephemeral tokens.
Bakit mahalaga ang Gemini 3.1 Flash Live sa Agosto 2026
Inilunsad ng Google ang Gemini 3.1 Flash Live sa Gemini Live API noong Marso 2026 — sa Agosto ito ang inirerekomendang model para sa production voice+vision agents. Pinapanatili ng Live ang persistent WebSocket session nang walang hiwalay na STT/TTS.
Agosto 2026: deprecated ang 2.5 Flash Live; mature na ecosystem; Live + Gemini 3.7 Flash (anunsyo ng 3.7 Flash).
Architecture: WebSocket sessions
Live API is stateful — persistent WSS bidirectional streams for audio, JPEG/PNG video (≤1 FPS), and text. First message: BidiGenerateContentSetup; then BidiGenerateContentRealtimeInput.
| Layer | Role |
|---|---|
| Transport | Stateful WSS |
| Input | PCM16 @ 16 kHz, video ≤1 FPS |
| Output | PCM @ 24 kHz, transcripts, tools |
Server-to-server (proxy) or client-to-server (ephemeral tokens required).
Native audio pipeline
Input PCM16 16 kHz; output 24 kHz. Browser float32 → PCM16. VAD for turns; manual activityStart/activityEnd for push-to-talk.
Camera and screen share
1 FPS JPEG/PNG — camera Q&A, screen code review. Default TURN_INCLUDES_AUDIO_ACTIVITY_AND_ALL_VIDEO — send frames only during speech to reduce cost. Limits: 15 min audio-only, 2 min audio+video.
Tools, thinkingLevel, security
Sync tool calling + Search grounding. thinkingLevel: minimal–high. 90+ languages. Ephemeral tokens — never embed API keys in clients.
Migration from 2.5
Use gemini-3.1-flash-live-preview; thinkingBudget → thinkingLevel; send_realtime_input for text; multi-part events; no NON_BLOCKING tools.
3.1 Live vs 3.7 REST
Live: WebSocket, sub-second voice, 128K context. 3.7: HTTP, coding/RAG, 2.5M tokens.