Timestamp generation for audio files (in all supported formats) is still broken for Gemini 2.0 production models (2.0 flash and 2.0-flash-lite), and can be verified in AI studio by anyone.
This issue has been brought up several times and there is still no acknowledgement from anyone at Google.
Right now the two 2.0 models in preview: Gemini 2.0 Flash (Image Generation) Experimental and Gemini 2.0 Flash Thinking Experimental 01-2 both produce accurate timestamps however they are obviously not suited for production.
Is anyone at Google aware of this issue and are there any plans to release a 2.0 GA model that produces accurate timestamps or will this pattern of taking away preview functionality upon general release continue?
Hey @rlev
I tried uploading a mp3 file in AI Studio and referred to a timestamp and it worked perfectly fine. Please let me know the file format used by you which broke the model.
Below are some of the observations.
The generated output is inaccurate and inconsistent in gemini-2.0-flash and gemini-2.0-flash-lite compared to gemini-2.5-pro-preview-03-25 and gemini-2.0-flash-exp-image-generation models.
Hi @Sangeetha_Jana, thank you for the response and looking into this issue.
To clarify: when I say that timestamp generation is “broken” what I mean is exactly what you observed: referring to timestamps is returning wildly inaccurate/inconsistent timestamps in generally available versions of the gemini 2.0 models (2.0-flash and 2.0-flash-lite).
Currently the only generally available model with accurate timestamp referral is gemini-1.5-flash-001 which will soon be retired.
I have tested mp3, AAC and WAV audio formats which all experience the same issue and I suspect it applies to all audio formats.
Right now it looks like this specific functionality is always degraded once the model goes from preview to GA. It would be nice to have some clarification whether this is considered a bug and are there any plans to fix this.
