Trying gemini 3.1 flash live and I can't seem to make it start talking?

Getting Started
1

I always have to be the one to talk first for some reason.

2

I think its the same problem Im having, I think its to do with the generate_reply(), on 2.5 live it was a universal command, but in 3.1 live, they have different commands for sending in the beginning of the call, and during. There are a bunch of migration issues which seem to be popping up and I really hope livekit fixes it.

Check out: https://ai.google.dev/gemini-api/docs/models/gemini-3.1-flash-live-preview

3

interesting! yea it would be great if they can look into that.

4

Thanks for raising, the team are aware.

I see this community-submitted fix: fix(google realtime): support gemini-3.1-flash-live-preview A2A model by youpesh · Pull Request #5238 · livekit/agents · GitHub , I suggest monitoring that for updates (it will either be merged, or closed as overcome by events if the issue is fixed by another PR)

5

Actually, this is probably a better PR to follow: fix(google realtime): support gemini-3.1-flash-live-preview by Hormold · Pull Request #5251 · livekit/agents · GitHub

6

the PRs are closed is livekit fixing this issue?

7

This PR is still open: generate_reply times out on first dialogue with gemini-3.1-flash-live-preview — works fine with gemini-2.5-flash- · Issue #5260 · livekit/agents · GitHub

I also gave an update in this other thread:

8

The matter remains unresolved; what timeline might we anticipate for its resolution?

9

@chitraksha_kaushik still an issue. Last I heard the team were working with Google on a proper fix, but let me check

10

While @darryncampbell checks the upstream timeline, the workaround that unblocks things today is to skip generate_reply() for the opener and use session.say() instead.

  class MyAgent(Agent):
      async def on_enter(self):
          await self.session.say("Hello, how can I help today?")

say() pushes a text turn directly for synthesis, bypassing the generate_reply() first-dialogue timeout in agents#5260 (generate_reply times out on first dialogue with gemini-3.1-flash-live-preview — works fine with gemini-2.5-flash- · Issue #5260 · livekit/agents · GitHub). For the rest of the session, generate_reply() works normally; the bug is specifically on the first dialogue per the issue.

Two other fallbacks:

  • Pin to gemini-2.5-flash-live-preview until Google ships the model-side fix.
  • Switch to openai.realtime.RealtimeModel if Gemini-specific features aren’t critical.
11

any updates regarding this issue?

12

@mayank_yadav, status unchanged since my last post. Issue #5260 is still open on livekit/agents, and no upstream fix from Google has landed. The session.say() workaround in my prior reply unblocks the opener if you need to ship in the meantime. For the upstream timeline, @darryncampbell was checking in #9; tagging him here in case he has a fresher read from Google.

13

Continuing the discussion from Trying gemini 3.1 flash live and I can’t seem to make it start talking?:

when can we expect this issueee to be solved? itss been 3 months!!! @livekitteams @darryncampbell @LiveKit-Community @livekit

14

I will check with the team to see if it is possible to fix. I don’t recall what the primary issue is with fixing this.

This doc points out that it is not currently possible and that a solution is being investigated

Last I checked, this was an issue that needed to be fixed on the Gemini side and not the LiveKit side:

If you believe this is wrong, please open a PR, and we will review it.