Is there a way to pass options to whisper?

[PacoH]

I miss the default --verbose True behavior of whisper. Is there a way to pass this and other options to whisper from subsai CLI? Maybe using a JSON string like '{"verbose": "True"}' as in the API?

[absadiki]

@PacoH, Yes, it is already the case. You can use a JSON string in the CLI, like the following:

subsai ./assets/test1.mp4 --model openai/whisper --model-configs '{"model_type": "small", "verbose": "True"}'  --format srt

Let me know if you find any issues with it ?

[absadiki]

Yes everything is documented on the API docs .. In order to know what configs to use you just need to check the config_schema variable of the target model, for example here are all the configs for the openai/whisper model.

For the output filename, I didn't implement the option to change the filename when using the CLI, because I thought it will be used just for batch processing .. But you can do it pretty easily using the web-UI and also using the package from Python.

You can use the following script for example:

from subsai import SubsAI

file = './assets/test1.mp4'
model_name = "openai/whisper"
model_type = "base"
subs_ai = SubsAI()
model = subs_ai.create_model(model_name, {"model_type": model_type})
subs = subs_ai.transcribe(file, model)
subs.save(f'test1-{model_name}-{model_type}.srt')

[PacoH]

OK, that's the same as in the API reference except for the 'f' before the filename in subs.save, which I assume is a typo. But how do I run this from the command line on OS X? I would use the API but I really can't find any useful information on how to implement it on OS X. I am only familiar with REST API usage.

[absadiki]

The f is not a typo, it is Python syntax for string formatting 😅

To run it from the command line, you can save the script in a file and call it with the python interpreter, there are plenty of resources out there how to run a python script from the command line.

But anyways, I added that option to the CLI for your convenience , you can now add a suffix to the output filename as follows:

subsai ./assets/test1.mp4 --model openai/whisper --model-configs '{"model_type": "small", "verbose": "True"}'  --format srt --output-suffix "openai-whisper-small" 

The output filename will be test1-openai-whisper-small.srt

Hope this helps!

[PacoH]

Wow, thank you for this addition.

First I upgraded openai-whisper and got:

pip3 install -U openai-whisper

Installing collected packages: openai-whisper
  Attempting uninstall: openai-whisper
    Found existing installation: openai-whisper 20230124
    Uninstalling openai-whisper-20230124:
      Successfully uninstalled openai-whisper-20230124
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
subsai 1.2.0 requires openai-whisper==20230124, but you have openai-whisper 20230918 which is incompatible.
Successfully installed openai-whisper-20230918

Then I tried to upgrade subsai and got:

pip3 install -U git+https://github.com/abdeladim-s/subsai

Installing collected packages: openai-whisper
  Attempting uninstall: openai-whisper
    Found existing installation: openai-whisper 20230918
    Uninstalling openai-whisper-20230918:
      Successfully uninstalled openai-whisper-20230918
Successfully installed openai-whisper-20230124

subsai --version
subsai 1.2.0

20230124 is a much older version than 20230918. Can you possibly update the tool to use the latest version?

[absadiki]

I can update the package to the latest openai-whisper version, but the problem is it will lead to so many conflicts because other packages (like whisper-timestamps and whisperX) depend on different versions of whisper, so if you update just openai-whisper, other packages won't work.
Besides, there are no major differences between the two versions I think, so I thought it is better to just stick with the actual version.