# machine-learning
- u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:37 PMno need to do that for taco - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:37 PMpipeline is proper multispeaker - e
Emoto
03/24/2022, 4:37 PMPipeline? - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:38 PM#841437191073955920 - e
Emoto
03/24/2022, 4:39 PMHuh is that a new notebook, because I remember I used a different one to train voices back in 2021 - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:39 PMyes - e
Emoto
03/24/2022, 4:39 PMThe one that used what's it called - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:39 PMits in development but it has a lot of cool things - e
Emoto
03/24/2022, 4:39 PMCool I'll check it out, thanks - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:39 PMlike proper multispeaker - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:39 PMgst’s - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:39 PMautomatic arpabet conversion - e
Emoto
03/24/2022, 4:39 PMI wasn't even aware like yoi guys updated that stuff still, and thought you kept using the old ljspeech tacotron model lol - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:40 PMits not fully rolled out yet - e
Emoto
03/24/2022, 4:40 PMWhile that one is good, it's kinda awkward at the same time as it's only good on female voices - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:40 PMthe site still uses legacy - e
Emoto
03/24/2022, 4:40 PMYeah it's still a wip isn't it? - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:40 PMyes - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:40 PMim currently working on training a massive multispeaker with 166 voices - e
Emoto
03/24/2022, 4:41 PMThat sounds amazing lol - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:41 PMsomeone else is already training a 53 voice multispeaker and even the unfinished product has done wonders with low wav datasets when warm started off of - e
Emoto
03/24/2022, 4:41 PMYeah that's expected with multispeaker datasets, so I'm glad it has the same effect on tacotron - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:42 PM(see: the morshu in #947423643229626409 ) - e
Emoto
03/24/2022, 4:42 PMAre the models still trained in 22k? - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:42 PMyes they are - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:42 PMbut 44k is possible i believe - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:42 PMand is being experimented with - e
Emoto
03/24/2022, 4:44 PMWas thinking of training a character voice in 44k eventually, once I have a full dataset (I'm planning on recording CMU artict + Havard sentences + maybe some of timid and ljspeech lol) - u
(Dawn) Will Draw Fictional Women
03/24/2022, 4:44 PMlegacy models are also supported for warm starting - d
Deagan
03/24/2022, 4:45 PMThank you for your insight, it means a lot! So TalkNET likely produces a lower quality end-result, when compared to Tacotron? How about FastSpeecher2? I would plan to train for very long periods, to ensure the best result. Could you link me to the latest training notebook, I would like to commit a few tests on a large sample!