Referring to timestamps (Audio Understanding) is returning wildly inaccurate/inconsistent timestamps for both 2.5 Flash/Pro models, applies to both VertexAI and Gemini developer API’s.
Performance degradation started a few hours ago. Has there been a checkpoint change?
For anyone having the same problem, you need to explicitly set audio_timestamp parameter to True when making requests:
config=GenerateContentConfig(audio_timestamp=True)
This still did not solve our issue. We are facing massive drop in timestamp quality since last 3-4 days. It was working perfectly fine for us before that. Did you do any other change?