Use some of the CLI arguments inside my python code

[hanserasmus]

Hi all

I am sorry if this is a stupid question. I cannot seem to find any documentation on how to do this, or understand from the source code how to implement what I want to achieve. I want to, amongst other things, limit the number of threads whisper is allowed to use, as it will be running on a server that shares resources with other applications. So there are a couple of things that is available on the CLI, but does not seem to be available to the transcribe function. If you take the example below:

model = whisper.load_model("large")
result = model.transcribe("audio.mp3")
print(result["text"])

I would like to change it so that:

• it limits the whisper application to 2 threads
• it needs to make use of a model directory where the models have been downloaded into already
• it needs to return not text, but rather returning srt and vtt.

Any possible examples I can look at to achieve what I want to achieve? TIA for anyone willing to help.

[ryanheise]

You can pass those arguments like this:

import whisper
from whisper.utils import get_writer

audio_path = "audio.mp3"
model_dir = "path/to/dir"
output_dir = "path/to/dir"

torch.set_num_threads(2)
model = whisper.load_module("large", download_root=model_dir)
result = model.transcribe(audio_path, word_timestamps=True)
srt_writer = get_writer("srt", output_dir)
vtt_writer = get_writer("vtt", output_dir)
srt_writer(result, audio_path)
vtt_writer(result, audio_path)

[ryanheise]

Can you try copying and pasting from my answer? Looks like you may have mistyped it.

[hanserasmus]

I am so sorry @ryanheise that was so stupid of me. But, looking at your answer again, I am not sure if the github page is not updating properly or what, because this is what I see on my side:

But later on you said you meant load_model.

So not sure what is going on but anyways, updated my code to now read model = whisper.load_model(("large", download_root=model_dir)) and all that pycharm is left complaining about is:

[hanserasmus]

Ah the single parentheses is the answer it seems.
model = whisper.load_model("large", download_root=model_dir)

[ryanheise]

My( key must have been over excited :-)

[hanserasmus]

Will see how this goes, thank you SO MUCH for your kindness and patience.

[phineas-pta]

see how whisper pass CLI args: https://github.com/openai/whisper/blob/main/whisper/transcribe.py#L511