[GH-ISSUE #1987] More information about the pretrains' dataset. #1342

New issue

Closed

opened 2026-10-20 10:08:14 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-20 10:08:14 +00:00

Owner

Originally created by @codename0og on GitHub (Apr 20, 2024).
Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1987

@RVC-Boss We already know that for the dataset, the vctk corpus ( 108 speaker version ) was used but how about the processing?
Was there anything applied?

denoising
peak / rms normalization
compression
how about the dynamic range?

I am asking because, as much as I've done tons of models I still can't quite find anything useful in that regard based on my trainings;

A) Is it better to limit the dynamic range of the dataset to the possible maximum ( without distortion introduction ofc )
B) Maintaining it somewhat natural ( slight peaks taming + slight compression to even stuff out and then -2 or -3 db general norm )
B) Taking care of the harsher peaks / peaks in general but leaving the dynamic range alone

What kind of approach you think would be suitable for your pretrains?
I'd really benefit from such information, and I am pretty sure some other more advanced users too.
Thank you in advance!

Originally created by @codename0og on GitHub (Apr 20, 2024). Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1987 @RVC-Boss We already know that for the dataset, the vctk corpus ( 108 speaker version ) was used but how about the processing? Was there anything applied? - denoising - peak / rms normalization - compression how about the dynamic range? I am asking because, as much as I've done tons of models I still can't quite find anything useful in that regard based on my trainings; A) Is it better to limit the dynamic range of the dataset to the possible maximum ( without distortion introduction ofc ) B) Maintaining it somewhat natural ( slight peaks taming + slight compression to even stuff out and then -2 or -3 db general norm ) B) Taking care of the harsher peaks / peaks in general but leaving the dynamic range alone What kind of approach you think would be suitable for your pretrains? I'd really benefit from such information, and I am pretty sure some other more advanced users too. Thank you in advance!

bsakaguchi

2026-10-20 10:08:14 +00:00

closed this issue
added the
help wanted

documentation

question
labels

bsakaguchi commented

2026-10-20 10:08:15 +00:00

Author

Owner

@SCRFilms commented on GitHub (Apr 24, 2024):

@RVC-Boss We already know that for the dataset, the vctk corpus ( 108 speaker version ) was used but how about the processing? Was there anything applied?

I just checked some samples in the vctk dataset and it's really bad.

-tons of mouth clicks
-loud mic noise
-low frequency rumbling noise (could be DC offset issue)
-lacks breathes sounds
-lacks pitch variations (speaker's pitches just sits about 110hz to 200hz)
-lacks higher harmonic details (causes it to have flipping harmonic and static harmonic artifacting)

I don't think they even apply processing to the audios, the dataset is also bad in the first place.

@SCRFilms commented on GitHub (Apr 24, 2024): > @RVC-Boss We already know that for the dataset, the vctk corpus ( 108 speaker version ) was used but how about the processing? Was there anything applied? I just checked some samples in the vctk dataset and it's really bad. -tons of mouth clicks -loud mic noise -low frequency rumbling noise (could be DC offset issue) -lacks breathes sounds -lacks pitch variations (speaker's pitches just sits about 110hz to 200hz) -lacks higher harmonic details (causes it to have flipping harmonic and static harmonic artifacting) I don't think they even apply processing to the audios, the dataset is also bad in the first place.