assuming i have 500 audio files

is there any like, telltale signs that i'm overtraining?

just asking because i want to leave it be for like 2-6 hours because i'm leaving and i can just turn on the notebook and leave it training

ofc i shouldnt have to worry about anything, i backed the model up

with big datasets people often take like 5% of the clips and don't use them for training, only for validation (the loss value) but the tacotron notebook doesn't have that because we usually don't have the data only other way i know of is to check whether it perfectly copies the intonation of your training data, but that requires switching between training and synthesis notebooks

yeah i'm just gonna leave it be and yolo

pf

i added a backup feature to my thing

i just backed up manually my 0.15 loss model

my audio clips arent perfectly snipped but it trains and sounds fine, so i assume that's ok

like sometimes words from the next clip get into the previous clip

and words from the previous clip get into the next clip

i think it's for about 20-30% of the whole dataset, but the text does match the usable data in the audio file

i mean it'll probably learn to synthesize those word bits every so often

also, can i just switch out the dataset with, say, another video, which would be another 400 completely different audios?

to change the dataset and continue training on new data basically

yeah absolutely

because i have multiple clean videos with auto subs and it doesn't take long to prepare them

idk why you wouldn't train on all 900 though

quality idk

for now i'm just using the 500 current files

more training data = more quality

it can probably only get to a certain point of quality anyway, beyond which it's diminishing returns

unless you have some noisy clips in which case you could try training on all then training some more on just the clean ones

i want to also try training on papanomaly, but he doesn't speak much in the videos, so it may take a while haha

i should be able to source at least 50-100 clips though

and he also has a twitch with VODs

no auto subs though, sad

Does the uberduck website use Tacotron 2 models or are they are modified version?

tacotron 2 + hifi-gan for the 📣 section, talknet for the 🎵