It seems Gemini 3 has lost its audio data interpretation capabilities, especially regarding time

Gemini API
1

When I input audio data into Gemini 3 to generate subtitles or timestamps, it produces nonsensical results. Gemini 2.5 Pro, however, outputs accurate data without any issues. I tested both models on AI Studio.

Is this a bug in Gemini 3, or an intentional feature downgrade?

3

Hi @KichangKim ,
Could you please share an example with steps to reproduce the issue, it helps us to reproduce and showcase to the appropriate teams.
Thanks.

4

It is simple. You can reproduce by these input.

  1. Upload arbitrary song MP3 audio file.
  2. Then, input this prompt:
    Please make the following lyrics into an ASS subtitles file exactly according to the time stamp of the attached MP3 file. Make sure the lyrics are marked in two lines at a time, but not across sections divided by . Designate style 1 for one-line subtitles. No comments.
    Lyrics’:
    [here is actual lyrics for song]
  3. You can see that Gemini 2.5 Pro makes precies ASS formatted text, and Gemini 3 Preview output incorrect time stamped ASS text.
6

@Nireeksha_K_A

I am experiencing similar difficulties when using Gemini 3 Pro for transcribing audio. Around a week ago everything was fine.

But today I tried to transcribe an audio file with a duration of slightly over 1 hour and there was very weird behavior occurring.

In the first test Gemini would continue transcribing same parts of the conversation over and over again (like 20+ times) until it ran out of the output context window.

The second time, it kept repeating his thinking process over and over again in slightly different ways (Kind of like the following: I’m now analyzing text, I’m now checking the text, I’m now analyzing, etc), without going to output the text (so I stopped him).

On other audio files there is (1) a very big degradation of quality, like enormous and (2) only completing very little of the total transcription when asked for a full transcription.

I can not share exact details as they are sensitive and I’ve used a Vertex project for this. I was using the Vertex AI Studio for this.

I hope it will get back to what it was about a week ago as it was working well for this exact use case…

1 Like