# machine-learning
- l
Lexi (delulu posts on the daily)
05/12/2022, 7:25 PM(so sorry for ping and for dumb question ) ok train for how long or until what number - d
Draconiator
05/12/2022, 7:50 PMIdeally you want to be at 0.012 or lower, and wait till your graph looks like a 45 degree diagonal line leading to the upper right - l
Lexi (delulu posts on the daily)
05/12/2022, 7:51 PMfor talknet - l
Lexi (delulu posts on the daily)
05/12/2022, 7:51 PMi was talking about talkne - l
Lexi (delulu posts on the daily)
05/12/2022, 7:51 PMt - d
Draconiator
05/12/2022, 7:51 PMAh, well that one I don't know. - h
hecko
05/12/2022, 7:54 PM0.12 not 0.012 - h
hecko
05/12/2022, 7:55 PMfor talknet you can stop when it starts saying "not in top 3" a bunch - l
Lexi (delulu posts on the daily)
05/12/2022, 7:56 PMok - k
Kelly
05/12/2022, 8:33 PMDoes anyone want to use my voice on uberduck? - u
{K EY1} (Kei)
05/12/2022, 8:34 PMWhat voice is it - d
dzlandis
05/13/2022, 1:00 AMGood or bad so far? - u
user
05/13/2022, 1:03 AMgood - d
dzlandis
05/13/2022, 1:12 AMI am getting a few like this though 🤔 - d
dzlandis
05/13/2022, 1:13 AMProbably okay since it's not fully trained yet I suppose 🤷♂️ - u
{K EY1} (Kei)
05/13/2022, 2:01 AM@dzlandis graphs with diagonal lines that just stop or go weird at a certain point are fine. It just means it's displaying the training of a shorter wav file The gaps, on the other hand, might be a tiny issue; they should clear up though - u
{K EY1} (Kei)
05/13/2022, 2:02 AMNp - u
user
05/13/2022, 8:48 AMcompare manual lip sync with ai - u
user
05/13/2022, 9:20 AM@Justin what do you think? - m
mildrefusal
05/13/2022, 11:56 AMI'm finding that, if a word isn't in that dataset, my model can't pronounce it at all. I thought, since it's using a pretrained model to begin with that it would be able to pronounce basically all words, just some wouldn't sound like your character? Is this not how it works I'm using the Pipeline btw, but I think this question applies to all - u
user
05/13/2022, 12:01 PMnot enough data in general - u
user
05/13/2022, 12:02 PMwhat voice are you trying to make - m
mildrefusal
05/13/2022, 12:02 PMSean Connery/James Bond. - u
user
05/13/2022, 12:02 PMsend the dataset - m
mildrefusal
05/13/2022, 12:03 PMI'll dm you - h
hecko
05/13/2022, 12:18 PMidk about words but it's definitely prone to forgetting phonemes because it has no incentive not to one thing you can do is train on multiple voices at a time in what's called a multispeaker model, then the voices can generalize from each other somewhat but of course that requires more data (but less per voice) - t
Toad
05/13/2022, 12:23 PMWhen I am inputting text if the AI can’t say a word I just spell the word phonetically - m
mildrefusal
05/13/2022, 1:20 PMIs there a general male voice dataset I could use that would cover the gaps in my own one? - h
hecko
05/13/2022, 1:21 PMmmm i was gonna say you could get some from
but the link for that is broken, lemme ask scruffy if he has a new oneaction 53