# machine-learning
- u
{K EY1} (Kei)
03/25/2022, 4:04 PMOh - u
{K EY1} (Kei)
03/25/2022, 4:04 PMWell What does it do? - u
(Dawn) Will Draw Fictional Women
03/25/2022, 4:05 PMit just splices them together - u
{K EY1} (Kei)
03/25/2022, 4:05 PMI would figure as much Are there any examples? - u
(Dawn) Will Draw Fictional Women
03/25/2022, 4:05 PMthe entire stupid “hybrid voices” category - u
{K EY1} (Kei)
03/25/2022, 4:06 PMI feel like i'm being judged - u
(Dawn) Will Draw Fictional Women
03/25/2022, 4:07 PMno i just hate hybrid voices - u
(Dawn) Will Draw Fictional Women
03/25/2022, 4:07 PMespecially being allowed on the site - u
Uberduck
03/25/2022, 4:08 PM@hecko quacked: Here's an example of a hybrid voice. As you can hear, it kinda sucks. - u
(Dawn) Will Draw Fictional Women
03/25/2022, 4:08 PMtakes every fiber of my being to not just deactivate them all and retire the category forever - u
(Dawn) Will Draw Fictional Women
03/25/2022, 4:08 PMseriously doubt anyone will miss them - h
hecko
03/25/2022, 4:09 PMif you hoped it'd make one coherent voice with characteristics of both the mixed ones then i have good news for you - u
(Dawn) Will Draw Fictional Women
03/25/2022, 4:09 PMwait for pipeline - h
hecko
03/25/2022, 4:10 PMyeah or try it right now with my undertrained horse model which i'll send a link to once my internet lets me - u
{K EY1} (Kei)
03/25/2022, 4:10 PMI'm probably going to start a project where me + anyone who wants to participate tries to voice various synthesis libraries I don't care for whatever character hybrid shenanigans are going on with uberduck, though Tbh i'm not incredibly interested in most uberduck models because they're just existing characters using random lines - h
hecko
03/25/2022, 4:10 PMoh so you wanna record your own with a custom reclist? nice - u
{K EY1} (Kei)
03/25/2022, 4:10 PMYeah - h
hecko
03/25/2022, 4:11 PMi've been thinking of making one that would include all the phonemes semi-proportionally (so the ai doesn't pronounce
asbeige
) along with maybe emotions?beize - h
hecko
03/25/2022, 4:12 PMbecause apparently there's no- actually hold on i wrote about this before - h
hecko
03/25/2022, 4:12 PM> because as far as i can tell there doesn't seem to be any single one that checks all the boxes of > - having multiple sentences in one line sometimes > - having question marks and exclamation marks > - not having obscure words > - representing all the phonemes in a low amount of lines > - (maybe) being of uniform length - h
hecko
03/25/2022, 4:14 PMthe ljspeech transcript file is good enough for wikipedia-style models but not emotional conversation-style ones but then again the current training notebook isn't good at those either - h
hecko
03/25/2022, 4:16 PManyway here's the speaker mixing notebook https://colab.research.google.com/drive/1wXWuhnw2pdfFy1L-pUzHfopW10W2GiJS - u
{K EY1} (Kei)
03/25/2022, 4:23 PMOh cools - s
Senator Armstrong
03/25/2022, 4:33 PMey - e
Emoto
03/25/2022, 4:59 PMI was thinking of getting that going too just to create a hybrid model for pretrained model reasons, but I'm waiting patiently for pipeline lmao - e
Emoto
03/25/2022, 5:02 PMThough, I wonder how pipeline been set up to allow voice merging. Cause that'd be interesting for pretrained model experiments actually - e
Emoto
03/25/2022, 5:02 PMWhere you'd get a bunch of speakers recording the same lines like CMU artic - e
Emoto
03/25/2022, 5:09 PMOh, the pipeline training notebook already allows for multispeaker models, so technically we could train multispeaker models?? - e
Emoto
03/25/2022, 5:09 PMThat we could use for like.. pretrained models?