# machine-learning
- s
SidPlays_144p
06/14/2021, 3:45 AM - m
M!nerman
06/14/2021, 3:46 AMI'd go for the movie version, if only because it's more iconic - m
M!nerman
06/14/2021, 3:47 AMThey're similar enough that you might be able to get away with combining the dataset - m
M!nerman
06/14/2021, 3:50 AMI'd train it on the game voice, if there's more data for that, then use that model as the base to train the movie voice on - m
M!nerman
06/14/2021, 3:50 AMIt should be noted, however, that I have no idea what the hell I'm talking about - w
Waltman13 (semi-active)
06/14/2021, 3:52 AMno I get ya. - w
Waltman13 (semi-active)
06/14/2021, 3:53 AMI only asked because I wanted to see if the stand-in actors were more safer to use. - v
VeryThickBoy
06/14/2021, 3:53 AMQuick question, how many Epoch's should I have to get a solid voice? - w
Waltman13 (semi-active)
06/14/2021, 3:55 AMI usually do 200. - v
VeryThickBoy
06/14/2021, 3:56 AMthe slope is getting up there - m
M!nerman
06/14/2021, 4:02 AMIs there any particular reason 22 KHz is necessary? - m
M!nerman
06/14/2021, 4:03 AMIf higher quality data is available, why not use it? - w
Waltman13 (semi-active)
06/14/2021, 4:07 AMNot sure honestly, it just do be like that I guess. - v
VeryThickBoy
06/14/2021, 4:08 AMif it isnt geting any closer to the redish orange data, should i just scrap it? - m
mega b
06/14/2021, 4:11 AMThere is not any particular reason. It may be because it is the minimal frequency we wont notice any bad audio, so it trains faster - i
IBob012
06/14/2021, 4:13 AMi remember pi saying that its the lowest it can be without we instantly noticing that its low quality + trains faster than full quality - m
mega b
06/14/2021, 4:13 AMyep - m
M!nerman
06/14/2021, 4:15 AMBut what if we need something even higher quality - w
Waltman13 (semi-active)
06/14/2021, 4:27 AMI guess I'll go with Spade. I mean it is the more recognizable voice, so muttering cause I don't know what else to add - u
_n
06/14/2021, 4:49 AMin every tutorial i've seen, the same dataset is used for training and validation. usually in machine learning the dataset is split in half (or some other percentage) and one half is used for training and the other is used for validation. - u
_n
06/14/2021, 4:49 AMwhy is that not the case with tactotron2? - i
IBob012
06/14/2021, 4:50 AMaccording to cookie validation list for tacotron 2 or something should be 5% of the list - i
IBob012
06/14/2021, 4:50 AMno clue why we dont do that - u
_n
06/14/2021, 4:56 AMi can kind of understand if the dataset is very small (less than 100 clips) but i'm surprised it's not really used or explained very well - u
_n
06/14/2021, 5:06 AMhow many steps/epochs do you typically do when training HiFi-GAN? - m
mega b
06/14/2021, 5:07 AMkeep an eye out on the alignment - i
IBob012
06/14/2021, 5:07 AMthat hifigan training colab doesnt show alignment for whatever reason (cringe) - u
_n
06/14/2021, 5:08 AMyeah so far im 90 epochs/2000 steps in - u
_n
06/14/2021, 5:09 AMthe notebook says a minimum of 5000 iterations are recommended (does that mean steps?) - i
IBob012
06/14/2021, 5:10 AMfor the amount of iterations i trained my taco model i train my hifi model for the same amount of steps