Some wired repetition happens in transcription.

[catalwaysright]

I am testing the medium.en model and I found this wired thing happens in my transcription.

There should be only one 'I am her mom'. Has someone met the same problem?

[jongwook]

This is a failure mode of the sequence-to-sequence architecture (briefly mentioned in the model card). Lowering the value of --compression_ratio_threshold (2.4 by default) might help.

[rolaguna]

Thanks! just want to add: in my case I was also setting temperature=0, so instead I let it as default values, decreased the compression_ratio_threshold as you suggested, and the problem seems fixed now

[bradyj04]

What did you end up lowering the threshold to? Im having the same issue

[fleek]

try changing "condition_on_previous_text" to false in transcribe and also lower the "no_speech_threshold"

[wiseman]

I would expect that setting condition_on_previous_text to False would cause an overall significantly decreased accuracy. Am I wrong?

[fleek]

I think it will affect if you are transcribing long narrated passages, especially if the sentence crosses the 30 second window. But my use is for conversations which are basically very short sentences, so I don't expect it to be a accuracy hit. Anyway, I ran the same file through whisper with the same settings 10 times and it never gave me the exact same output twice, so I cannot gauge whether there is any accurarcy hit or not.

[nlgtuankiet]

Hi @jongwook do you have any comments regarding the accuracy of the conversation use case? (< 30 seconds sentence)

[RichardQin1]

My audio usually lasts for 30 min. Is this suggestion also applicable. What are the recommended values for setting？thank