# machine-learning
- m
Mitchell
08/25/2021, 3:15 AMthis @User - h
Holos
08/25/2021, 3:57 AMI just deleted the ones that had to be downsampled unfortunately. I think out of the whole pony dataset i'll only keep the ponies that are longer than a few hundred files - t
the knight of the flowers
08/25/2021, 9:43 AMHowdy everyone. Real quick question in regards to Google Colab's GPU: if I were to switch accounts would I receive a new GPU? I've been working on training this one model with 292 wavs, but I had to redo it entirely because I believe I overtrained it. And now I've been getting the K80, even after repeatedly factory resetting runtime and holding off on using Colab. - h
Hugh G. Rection
08/25/2021, 11:54 AMAnother new voice model in the works: - u
user
08/25/2021, 1:05 PM@UserIt finished successfully! - t
The Wiggles and High School D×D
08/25/2021, 1:06 PMGreat - u
user
08/25/2021, 1:07 PMYou would receive a new GPU but there's no guarantee that it wouldn't be a K80. However, I have had luck with that strategy. - t
The Wiggles and High School D×D
08/25/2021, 1:08 PMThe next step is weird Apparently it requires you to use the pretrained model and run with that, although the zip file looks pretty massive for only a small portion of it needed for an extraction - t
The Wiggles and High School D×D
08/25/2021, 1:08 PMI'm going to send something for the third step - u
user
08/25/2021, 1:09 PMSo which file is the successful file? The largest file I see in in the logs folder titled "my_model_loss_0.53_per_0.81-342000.meta", though it could be another file. - t
The Wiggles and High School D×D
08/25/2021, 1:34 PMI can't tell It might have been present when we first downloaded the code, but it doesn't appear in the GitHub repository. - t
The Wiggles and High School D×D
08/25/2021, 1:37 PMGo for the most recent files - t
The Wiggles and High School D×D
08/25/2021, 1:41 PMAll three file types of the most recent name created, plus the binary checkpoint, four files total - u
user
08/25/2021, 1:57 PMok - u
user
08/25/2021, 2:11 PMHoney from HouseBroken - b
BrandonTD
08/25/2021, 2:18 PMHow many samples do y'all think 15.ai's spongebob model had? - a
Ask for dataset = $25/hour
08/25/2021, 3:21 PMWhat's the current meta for automated transcription? - s
SBSPfan2089
08/25/2021, 3:31 PMif you submit a dataset, is it put on the website as a voice? - h
hecko
08/25/2021, 3:43 PM8 minutes ≈ 100 samples, according to this screenshot https://images.squarespace-cdn.com/content/v1/5c8d4324fb22a57f65b403f3/1589158258406-5QZQ2LH6PXUIYWOMWEPR/spongebobsquarepants%21?format=500w - b
BrandonTD
08/25/2021, 3:44 PMPerfect, thank you - h
hecko
08/25/2021, 3:45 PMthough note that 15 uses one neural network for all the voices (multispeaker model), so it needs less data per voice for the same quality (and it can borrow emotions from other voices which is the main selling point imo) - s
SBSPfan2089
08/25/2021, 3:57 PMKen someone help me? - s
SBSPfan2089
08/25/2021, 4:02 PMI cant make a dataset - h
hecko
08/25/2021, 4:10 PMwhat's the issue - s
SBSPfan2089
08/25/2021, 4:14 PMI'm blind and I can't create a dataset for uberduck. - u
user
08/25/2021, 4:14 PMIf you are blind, then how are you typing these words? - h
Holos
08/25/2021, 4:15 PM - h
Holos
08/25/2021, 4:15 PMApparently they made a really really good alignment - s
SBSPfan2089
08/25/2021, 4:20 PMI am typing dis on my phone with voiceover. - h
hecko
08/25/2021, 4:29 PMoh m