adjust the sentence length and the way how the model will split sentences

[AIHIsora]

Thanks for the great work! It's the most accurate tools to transcribe audios as far as I known.

But in my test I met with a problem that the model will split sentences in some unreasonable ways. Sometimes the clips are split in the middle of a sentence, and sometimes one clip contain a long blank between two sentences. Is there a parameter to adjust it to get better segment performance?
I'm now testing it with Japanese, if it matters.

[abdeladim-s]

Hi @AIHIsora,

You are welcome. 'Glad you found it useful.

Could you please elaborate more ? what model are you using ? have you tried other media files (other than Japanese) ?
Have you tried the provided VAD example:

• VAD example: process long audio files using silero-vad.

In general, you can get better performance by using a bigger model size!

[AIHIsora]

I was using whisper medium model (limited by my gpu memory), and only test in Japanese. I found the problem may be due to my test media, which is a mix of songs and talks that will contaminate each other. The VAD example seems to be a suitable solution, or maybe I can split the song and talk chunks manually.
Thank you for the reply. I will test further and will come back with feedback if I make progress or encounter new problems.