Feature Request: Adding ability to specify compute_type in transcriber section

[PsychoThinker]

Is your feature request related to a problem? Please describe.
I would like to use the default faster-whisper-large-v3-turbo-ct2 model with my GPU. However, my GPU does not support float16 compute types, so the compute_type would need to be set to float32. Currently, it seems there's no way to adjust this parameter in the configuration.

Describe the solution you'd like
I propose adding a compute_type option to the configuration file under the transcriber section, with a default value (e.g., float16). This would allow users to customize the compute type to match their hardware capabilities, such as switching to float32 when necessary.

[jonathanfox5]

Thanks, I'll take a look into this.

Out of interest, what GPU are you using?

[PsychoThinker]

Thanks! I'm using NVIDIA Pascal architecture (1080 ti).

There are other important parameters for a faster whisper model like: beam_size, condition_on_previous_text and batch_size but I can live without them :)

[jonathanfox5]

Thanks :) I might see if I can automatically set it based upon the detected GPU in a reliable way, otherwise I'll add an option.

I purposefully omitted the other factors for two reasons:

1. The gogadget transcribe built in help page was already pretty long and I was already quite worried about overwhelming less tech savvy users with lots of information.
2. WhisperX (which I'm using as the backend to work with faster whisper) is very buggy. I therefore didn't want to add in the extra work of testing and fixing the options across the 5 different platform configurations.

HOWEVER, I could expose all of the options through an --advanced-options flag like I did for gogadget download. Would this work for you?

[PsychoThinker]

Of course - I fully understand your concern about creating an application that is too complicated to use. Also, I don't think changing it will make that much of a difference when transcribing a video of about an hour's length. This "--advanced-options" flag would be nice to have option, but it seems to me that it would require an example in the documentation, because it would need to be somehow converted from string to list of parameters for python method.

I'm not a big fan of whisperX, if need to have super optimal solution I would use whisperS2T, but both of these library have not been updated recently compared to the original faster whisper.