Model transformation.

[boykan313]

Hello! I have a problem converting the whisper model to faster-whisper. After this process, I don’t understand what to do, the process is very difficult. Help me please.
I am using this model: https://huggingface.co/mitchelldehaven/whisper-large-v2-ru
Transformed through this command: ct2-transformers-converter --model mitchelldehaven/whisper-large-v2-ru --output_dir whisper-large-v2-ru-ct2 --copy_files tokenizer.json --quantization float16

[boykan313]

Now I'm learning such a result when transforming the model.

[guillaumekln]

Try converting without --copy_files tokenizer.json.

[boykan313]

I already understood this... I found the directory of this model, renamed it to large-v2 and launched it, downloaded one file and started the process. it seems to work ... I thought this model would be read as a separate one, but in the end I had to crush the existing model of this one, under the same name.
this process was so primitive?

[boykan313]

what you said worked only after the complete removal of the downloaded model, but without the command that you called. I'm glad it all worked out in the end. Thank you for your help and experience! Don't abandon this project!

[boykan313]

Tell me please, is it possible to somehow spoil the model during education? My new model, which is aimed at the Russian language, works worse than your universal one.

[boykan313]

Sorry for bothering you often, can you help me with this problem with my model? She does not want to start, whisper does not see her.

[guillaumekln]

It seems you are using https://github.com/Purfview/whisper-standalone-win to run the model, so you should ask questions in this other project. Your latest error has nothing to do with faster-whisper.

[Purfview]

I already understood this... I found the directory of this model, renamed it to large-v2 and launched it, downloaded one file and started the process. it seems to work ... I thought this model would be read as a separate one, but in the end I had to crush the existing model of this one, under the same name. this process was so primitive?

Don't replace existing model, just run converted model from new folder.

[boykan313]

figured it out! The name after the model conversion did not completely change the name of the folder, it was just a whisper, added faster and it all worked! Thank you for your patience and help! Develop the project further, I will follow the updates!

[Purfview]

Tell me please, is it possible to somehow spoil the model during education? My new model, which is aimed at the Russian language, works worse than your universal one.

It's possible.

[boykan313]

I did that and the transcription was terrible.

[Purfview]

Downloading same file wouldn't change your results.

[boykan313]

It turns out that this file is the same for all models? there is only a binding to medium, large, etc.?

[Purfview]

Yes.

[boykan313]

which is strange, running it in a normal whisker, it worked, but for some reason only a couple of seconds, then it freezes but the processor still loads.

[boykan313]

No.

The script created the directory whisper-large-v2-ru-ct2 so the conversion worked.

The script created the directory whisper-large-v2-ru-ct2 so the conversion worked.

The script created the directory whisper-large-v2-ru-ct2 so the conversion worked.

As I understand it, whisper does not see the model, because it does not contain the "tokenizer.json" file. In the instructions for transforming the model, it was said that this file should be downloaded later on the next start, but this does not happen. You can copy this file from the original large-v2 model and then the new model will be detected, but it will not work correctly..

[boykan313]

Sorry, I want to ask about the model transformation process. The transformation was successful but without the "tokenizer.json" file. Where can I get it now so that the model works? Your instructions say that it will download itself after the next launch, but this did not happen.

[Purfview]

I don't know much about the model transformation process nor if "tokenizer.json" file is mandatory.
Currently Standalone Faster-Whisper. checks if all 4 files are present, same named files as in guillaumekln's original models.

EDIT:
Try to copy tokenizer.json from some original model. [it's same in all original models]