Best results are with "word_level=True". But problem with translating the srt output

[dgoryeo]

Hi,

When using "word_level=True" and "task-transcribe" one gets most accurate results. However tranlating the srt output is a dilemma, because each line is repeated several times. Most of paid translation APIs are based on number of the words hence the cost would be much higher. Is there a way to be able to get around that somehow?

Thanks

[jianfch]

If by "repeating", you mean the line repeats but highlight a different word each time. Then that's the expected behavior.
If the lines are repeated in result, then try this openai/whisper#192 (comment).
Keep in mind that the word timestamps are not meant for translations. They are not reliable for translations.

[dgoryeo]

Thanks @jianfch , yes I do expect that behaviour and I do want that behaviour since that produces the most accurate results, in my experience.

The dilemma I'm facing is that: after the accurate subs are produced through transcription, I need to translate them to English --I use DeepL for that. The cost and time are a factor higher because I end up paying for the "word highlights". I am wondering if there is any way to avoid that. I'm guessing a possible solution would be to translate the results in the json file, then output the subs file. But my scripting skills is poor to be able to code that.

I'm not sure if anyone else has similar dilemma as I though.

[jianfch]

If you want to keep the highlighting but use less overall text, try to use vtt instead of srt.
Note that word_level is only for highlighting in the output file, it doesn't affect the accuracy of the results. The timing of the segments stays the same without the highlighting.