Gemini Live Model Regression

Gemini API
,
1

Hi everyone,

I’ve recently noticed that audio responses from the Gemini Live model appear less intelligent or precise than before (e.g. Mar 27th). Specifically, I’ve observed:

  • The model speaks unusually slowly for the first few words of its responses.
  • When ending conversations, it occasionally produces unnatural phrases, such as saying “You single quote re welcome” instead of “You’re welcome.”

Has anyone else experienced similar issues lately, or is this possibly related to a recent update or bug?

Thanks for your insights!

1 Like
2

I’ve had somewhat similar issues with multimodal Live API (audio output) and gemini-2.0-flash-exp, where there is slowness or pauses in the response. I believe things were working fine on 24 Mar 2025 and noticed there was an issue today.

1 Like
5

Just an update, I think today (April 1st)'s Gemini model is performing normally again

1 Like