Improve suggestion
#2
by
LukeJacob2023
- opened
Hello, antony66, thanks for your contribute of this model. I am trying to improve russian whisper too. In my option, finetune on common voice will not improve much, because I think the original whisper have collect the dataset and train. I think we need some accurate dataset. Generally after finetune, the wer can reduce 50%.
This model improve much in it's dataset. I think original whisper may has not train it.
https://github.com/sovse/base_rus_whisper_stt
And original whisper may not use dataset:
https://github.com/snakers4/open_stt
Hey! Yes this is exactly my thought as well. So I had been working on a custom dataset until I had to switch to another task temporarily. Looking forward to returning to this task soon
This comment has been hidden
LukeJacob2023
changed discussion status to
closed
LukeJacob2023
changed discussion status to
open