As users have previously pointed out, audio timestamp accuracy for the Gemini 2.0 flash lite model has been broken after going from gemini-2.0-flash-lite-preview-02-05 to GA and it seems to completely hallucinate them.
Something interesting I have found is if you take the same audio clip (tested with mp3 files) and convert it to a video format (tested with mp4 files with a solid background) the timestamps are accurate.
This is not an ideal workaround since it comes with a major increase in input tokens.
Are there any plans to fix this bug?