Why does it take so long to process a 1 minute video?

[Root-FTW]

How can I make it run faster? It takes too long for a 1 minute video.

I am using this CLI command:
whisper_timestamped --accurate video.mp4 --model large-v1 --output_format srt --vad False --device "cuda:0" --output_dir .

My PC:

Windows 11 PRO
GPU: GTX 1080 TI
CPU: i9-9900k

[Jeronymous]

How much does it take? dozens of minutes?

That is weird. Is there a chance you can share the video?
Is the transcription (almost) correct?

[IntendedConsequence]

Something doesn't feel right. When I run with --model large-v3, VRAM shoots to 10GB and it's all very slow. And yet, running the large-v3 model in whisper.cpp, and using vanilla transformers pipeline (in this notebook https://huggingface.co/spaces/hf-audio/whisper-large-v3/blob/main/whisper_notebook.ipynb), both of them only use about 4GB VRAM, and are noticeably faster.

[Jeronymous]

Thanks for spotting @IntendedConsequence

Unfortunately the issue of high VRAM consumption happens in the openai-whisper package itself 😬
openai/whisper#1670 here it's reported from version 20230918, but I tried previous versions (up to 20230124) and experience the same: openai-whisper always hits ~10GB VRAM for large models.