we need to load model once, and keep it there

[falves1]

I have a CUDA card with 12 GB of RAM; the large-v3 model takes 10 GB. With the command line command, whisper reloads the model in full every time you transcribe even a small file. If I use Python, the behavior is much more efficient: the model can be loaded once and pinned to the GPU. I need to achieve that with the command line version. How does it get done?

[glangford]

You can list multiple file names as the first argument(s) on the whisper command line, or use wildcarding (eg. *.mp3). whisper will load the model once and iterate through the set of files.

[glangford]

whisper/whisper/transcribe.py

Line 595 in ba3f3cd

for audio_path in args.pop("audio"):

[falves1]

will that work in plain bash? like whisper *.wav ?
how do I do monitor a directory so any files copies to it are transcribed without reloading the model, ever? Not using Pyton.