Best settings for Purfview Faster Whisper XXL using large-v3 (transcribing to .srt)?

[Neobond]

Hi guys, I am new to this speech to text thing in Subtitle Edit, I am just wondering if there are any other switches people could help me with to improve the accuracy and output that people can share?

My rig is 9950X3D with 5090FE and 64GB DDR5 6000MT/s.

Appreciated!

[darnn]

I can only speak for myself (I have a mere GeForce GTX 1660 SUPER), but the default settings work best for me. Also, I'd use v2 and not v3, unless you want to use the smaller turbo models. I've gotten slightly better results with v2 in English and the other two languages I've worked with.

[darnn]

I don't think I ever really tried --sentence since it would mess with the accuracy of the timing. I haven't messed with --compute_type since the default is auto, and I assume it chooses the best one my GPU supports, so I can only do worse from there. --beam_size 5 is already the default in Faster-Whisper-XXL. I've tried different temperatures and in the end the transcription is most accurate with the default. Occasionally the timing is more accurate for a specific segment with a different temperature, but it was never worth the effort to test them all for every new recording.

[Neobond]

IDK without these switches the FairlyOddParents theme song is way off, now it is mostly accurate, and even has a musical note symbol where it is sung.

[darnn]

Hey, if it works for you, knock yourself out. I'm also using v2, maybe that contributes to the difference. This is not directly related to Faster-Whisper-XXL, but you may also want to use a vocal remover to isolate the vocals before you use it. It sometimes swallows some dialogue, but gives better timing overall.

[Neobond]

How do I add a vocal remover?

[darnn]

Either through Faster-Whisper-XXL itself (add "--ff_vocal_extract mdx_kim2" when you run it), or through
https://ultimatevocalremover.com/
(I prefer Kim Vocal 1 myself).