Limit transcription CPU threads #5595

New issue

Open

opened 2026-02-22 11:46:22 -05:00 by deekerman · 3 comments

deekerman commented

2026-02-22 11:46:22 -05:00

Owner

Originally created by @JohnXLivingston on GitHub (Jul 5, 2024).

Describe the problem to be solved

v6.2.0-RC1 comes with an incredible feature: automatic subtitles generation.

But the models that are used can use a lot of CPU and RAM.
I did not see any option to limit their usage (as we can with video transcoding).

Describe the solution you would like

Would it be possible to add some options?

Originally created by @JohnXLivingston on GitHub (Jul 5, 2024). ### Describe the problem to be solved v6.2.0-RC1 comes with an incredible feature: automatic subtitles generation. But the models that are used can use a lot of CPU and RAM. I did not see any option to limit their usage (as we can with video transcoding). ### Describe the solution you would like Would it be possible to add some options?

deekerman added the

Type: Feature Request ✨

Component: Subtitles 💬

Component: Runners

labels

2026-02-22 11:46:22 -05:00

deekerman commented

2026-02-22 11:46:30 -05:00

Author

Owner

@Chocobozzz commented on GitHub (Jul 10, 2024):

I don't think we can limit RAM but CPU yes there is a threads option

  --threads THREADS     number of threads used for CPU inference (default: 0)

@Chocobozzz commented on GitHub (Jul 10, 2024): I don't think we can limit RAM but CPU yes there is a `threads` option ``` --threads THREADS number of threads used for CPU inference (default: 0) ```

deekerman commented

2026-02-22 11:46:30 -05:00

Author

Owner

@lutangar commented on GitHub (Jul 18, 2024):

RAM usage largely depends on model size since it must be loaded to RAM (multiplied by the number of runners since models aren't shared in memory).

For example with whisper-ctranslate2 (which uses faster-whisper which is CPU friendly) the models tend to be larger than the one provided for openai-whisper :

large-v3 ~3GB
medium ~1.5GB
tiny ~75MB

Of course, transcript quality will get worse the further you decrease the size.

https://huggingface.co/Systran

@lutangar commented on GitHub (Jul 18, 2024): RAM usage largely depends on model size since it must be loaded to RAM (multiplied by the number of runners since models aren't shared in memory). For example with `whisper-ctranslate2` (which uses `faster-whisper` which is CPU friendly) the models tend to be larger than the one provided for `openai-whisper` : - `large-v3` ~3GB - `medium` ~1.5GB - `tiny` ~75MB Of course, transcript quality will get worse the further you decrease the size. > https://huggingface.co/Systran

deekerman commented

2026-02-22 11:46:30 -05:00

Author

Owner

@JohnXLivingston commented on GitHub (Jul 18, 2024):

Thanks for the information @lutangar ! Good to know.

@JohnXLivingston commented on GitHub (Jul 18, 2024): Thanks for the information @lutangar ! Good to know.