Gemini3.1 live preview model drops mid conversation for odd seconds

Agents
,
1

Using gemini-3.1-flash-live-preview with LiveKit Agents and seeing occasional realtime audio stalls/truncated responses during phone calls.

Example:

  • Agent starts response: "للتأكيد، آخر"

  • Then ~19s silence

  • User starts saying “ألو”

  • Agent later continues normally

Logs also show:

  • received server content but no active generation

  • server cancelled tool calls

Model warning:
gemini-3.1-flash-live-preview has limited mid-session update support

Has anyone else seen:

  • partial audio generations,

  • delayed continuation,

  • or realtime stalls with Gemini Live Preview models?

Using:

  • livekit-agents 1.5.8

  • livekit.plugins.google 1.5.8

  • SIP outbound calls - self hosted

  • Arabic realtime voice conversations

2

@pddotabdur, the warning is verified in livekit-plugins google/.../realtime/realtime_api.py:295:

if “3.1” in model: logger.warning(f"'{model}' has limited mid-session update support. instructions, chat context, and tool updates will not be applied until the next session.`

The two log lines you see (received server content but no active generation at line 1143, server cancelled tool calls at line 1191) are symptoms of that immutability colliding with normal mid-call state changes. Same upstream Google-side issue as livekit/agents#5260 (open).

Two paths today:

  1. Swap to a non-3.1 Live model: e.g. gemini-2.5-flash-native-audio-preview-12-2025 (in KNOWN_GEMINI_API_MODELS, not gated by mutable = "3.1" not in model at line 271), or the prior gemini-2.5-flash-live-preview.
  2. Switch realtime provider: openai.realtime.RealtimeModel if Gemini-specific features aren’t load-bearing.

No client-side workaround for 3.1’s immutability; the plugin itself acknowledges it.