# machine-learning
- h
HolyArapaima
09/28/2022, 4:54 PMMaybe if you had a ton of amazing data and a lot of luck it would be more worth while. I have actually been currently looking at other tools and things to raise that synthesis quality and support reference audio. - u
{K EY1} (Kei)
09/28/2022, 4:55 PM@hecko you should send an example of csd It sounded quite good to me - h
hecko
09/28/2022, 4:56 PM - u
{K EY1} (Kei)
09/28/2022, 4:56 PMTmk you just need a lot of data. I tried with 15 minutes and it sounded like shite - h
hecko
09/28/2022, 4:57 PMyou might be interested in nnsvs, it needs more data prep but sounds much better - u
{K EY1} (Kei)
09/28/2022, 4:57 PMI am making a new talknet pretrained model though so that should work better with singing than the current pretrained model - p
PixPrucer
09/28/2022, 4:57 PMStill no other language support - p
PixPrucer
09/28/2022, 4:57 PM💔 - u
{K EY1} (Kei)
09/28/2022, 4:57 PM^^ You don't even need a lot of data for nnsvs. Minimum 15 minutes for best results, but if you train on a pretrained model you can get good results with 1 minute of data - u
{K EY1} (Kei)
09/28/2022, 4:58 PMYeah i can't figure out how to do that Neutrogic tried and failed cuz apparently it's really hard - h
HolyArapaima
09/28/2022, 4:58 PM^ - u
{K EY1} (Kei)
09/28/2022, 4:58 PMYou'll hear a lot of people say you need hours of data btw but as a low data supremacist you do not - p
PixPrucer
09/28/2022, 4:58 PMOh hi my work - h
HolyArapaima
09/28/2022, 5:00 PMI gave Neutrogic my singing dataset for talknet testing as I had it all prepped, I am actually recording another really big English dataset for someone as there isn't a lot of male singing datasets. - h
HolyArapaima
09/28/2022, 5:00 PMThere even gonna make a NNSVS of it - u
{K EY1} (Kei)
09/28/2022, 5:00 PMWould you be okay with contributing your singing dataset for the talknet pretrained model? There's female singing in it rn but no male singing - h
HolyArapaima
09/28/2022, 5:01 PMI will have to see but I may be able to do that - p
PixPrucer
09/28/2022, 5:01 PMReminds me of my singing Talknet dataset It's very whack though - u
{K EY1} (Kei)
09/28/2022, 5:01 PMWhack how? - u
{K EY1} (Kei)
09/28/2022, 5:02 PMCuz if it's ok (transcriptions are correct) Could i hab it for pretrained model 👀 - h
hecko
09/28/2022, 5:03 PMtemptation to yoink some vocals from ccmixter and do nnsvs to them - u
{K EY1} (Kei)
09/28/2022, 5:03 PMI'm making csd english nnsvs - p
PixPrucer
09/28/2022, 5:03 PMIt's stuck ok my PC at home o(-( Remind me to send it this Friday - u
{K EY1} (Kei)
09/28/2022, 5:03 PMOki - h
hecko
09/28/2022, 5:04 PM - h
HolyArapaima
09/28/2022, 5:05 PMI am making another beta big al model and it's getting a lot better I have been able to get crust to interpret a lot of the Russian data better it's actually training correctly with only the best and original data. - u
{K EY1} (Kei)
09/28/2022, 5:05 PMOoo - h
HolyArapaima
09/28/2022, 5:07 PMBut I have it planned for something with higher quality synthesis that can utilize reference audio so I get to work on it with someone today. - p
postmates!!
09/28/2022, 5:08 PMtraining my spamton dataset rn gonna see if its shit or not