Recitation Error Issue w/ Audio

Gemini API
,
1

I am using the Gemini API to generate transcripts for publicly available audio (no copyrights). Up until this point - no issues across a large dataset. However - for a given audio file, I am getting persistent RECITATION error reason issues (multiple retries). If the audio is transcribed elsewhere on the internet - it does not seem logical for the model to continue to reject the output if a similar transcription appears in the training set.

Anybody else having a similar issue, and for anyone from Google - any best practice on how to communicate to the model that this is a common audio file?

2

Welcome to the forum. I got RECITATION blocked when using this sample audio file: Arthur the Rat – Dictionary of American Regional English – UW–Madison
It used to be a somewhat useful test input to help count audio transcription errors, since the dataset comes with known good text transcription files that match the audio. But then one day, it stopped being accepted.

1 Like