Faster-Whisper-XXL Pro Linux with AMD GPU

[IlDavid86]

Hi everyone, i'll ask here after a confusing research online
I just upgraded to Pro anche changed the vad_method to Ten(default so i should delete the command?) and the vocal extract to mb-reformer, the command now is like this, any improvement is appreciated :-P
-l ja -m large-v2 --vad_method ten --ff_vocal_extract mb-roformer --task transcribe --standard -o source

Before, with silero_v4_fw and --ff_mdx_kim2, i have been using it on windows CPU only (i have a 7900XTX) and the time spent was not so bad
Now i'd like to to try with fedora 43 that i have in dual boot to utilize the GPU but i have no clue

Thank you for any advice :-)

[Purfview]

the vad_method to Ten(default so i should delete the command?)

Yes. Default means that it's used without you setting it.
"--task transcribe" is default too.

i have been using it on windows CPU only (i have a 7900XTX) and the time spent was not so bad

If your CPU is high end then speed is bearable, but expect mb-roformer to be much slower than mdx.

Now i'd like to to try with fedora 43 that i have in dual boot to utilize the GPU but i have no clue

Linux will not make your AMD GPU to have CUDA. 😁
It will still run on CPU.

the command now is like this, any improvement is appreciated :-P
-l ja -m large-v2 --vad_method ten --ff_vocal_extract mb-roformer --task transcribe --standard -o source

--vad_method pyannote_v3 is still the best for quality (but it's slowest)
Use --realign if you care about timestamps

[IlDavid86]

Thank for the answer Purfview

Linux will not make your AMD GPU to have CUDA. 😁
It will still run on CPU.

I read something about ctranslate2-rocm but i don't know how viable It is

[Purfview]

I read something about ctranslate2-rocm but i don't know how viable It is

#379 (comment)

[IlDavid86]

Well, that's It, CPU It Is 😁