Gemini 2.0 Flash and Gemini 2.0 Flash Lite completely hallucinates Timestamps when Transcribing audio

[bigmandevs]

Description of the bug:

The GA versions of both models completely hallucinate timestamps when performing transcriptions on Audio.

Actual vs expected behavior:

The timestamps should be accurate based on when that word or phrase was spoken. The preview models for both where excellent at this. The same models after going into GA completely hallucinate timestamps.

Any other information you'd like to share?

Interestingly, this does not apply to extracting timestamps from video, and is applicable to audio only.

You can replicate this by trying any audio file and comparing the accuracy to videos.

[Konikz]

Hi, I'm new to open source and interested in looking into this. Just wanted to clarify—does this issue come from the JavaScript SDK, or is it a backend model problem? If there’s anything that can be fixed at the SDK level, I’d love to help.

[shmishra99]

Hi @bigmandevs ,

I also tried the sample audio with the JS SDK and Gemini 2.0 Flash, and it generated an incorrect timestamp. However, when I tried the same audio with Gemini 2.0 Flash in Google AI Studio, it generated a correct timestamp. This seems to be an issue with the JS SDK, not the model. I will dig into this issue further and provide you with an update.

Thank You!!