i mean 80 epoch

this good?

im missing a 2 because i deleted it for being too short

that's gonna break the checker i'm pretty sure. i just had an issue like that earlier

it shouldn't matter actually

i don't even do numbers in my datasets, i just use whatever filenames the files already have

you actually don't want to bring the error too low down with such a small dataset, because it'll start memorizing the lines and fragments of the lines and eventually it'll forget how english works due to obsessing over the lines

that sounds a lot more sane. what i get for slavishly following the guide

that sounds exactly like what just happened.

😄

yeah

i've had that happen to a lesser extent with my seal dataset

where it just decided that if the word

this

is at the end of a sentence then it must be pronounced with a high pitch

i was trying the colab voice clone that could do one 5-12 second file and then you just manually reload it and it always wanted to hit the word 'absorb' with a super high pitch.

i have the 5-second clone thing on my computer and it kinda sucks

i mean for 5 seconds it's good but ehhh

especially since it's trained on normal people voices not character ones

right - for me it began to get tantalizingly close and then would freak out and breathe and squeak for andother 12 iterations. interesting that you have the local one. i only could use the colab one due to an error installing numba on my Mac. it doesn't have NVIDIA though so i don't know if i could even run it.

it failed on numba while trying to install pytorch

anyway for the model with limited files - so 100 is okay? and do you just let it complete the epoch count or is there a good error number to shoot for?

Random, but is TalkNet only for the singing voices section or is it also for the normal section?

unfortunately only the singing voice section so go bug zwf to change that

unrelated if anyone ever wants to use the /mlp/ 4chan ponytools pony ppp datasets here's a script to convert them to the format our notebooks use

py
import os
import shutil
import json

download_from_google_drive("[drive id here]", "dl.tar")
os.mkdir("/content/tacotron2/wavs2")
!tar -xvf dl.tar -C /content/tacotron2/wavs2
os.remove("dl.tar")

listtxt = []
for dirname, dirnames, filenames in os.walk("/content/tacotron2/wavs2"):
  for i in dirnames:
    with open(os.path.join(dirname, i, "label.json")) as f:
      data = json.load(f)
      listtxt.append("wavs/" + i + ".wav|" + data["utterance"]["content"])
      os.rename(os.path.join(dirname, i, "audio.wav"), os.path.join("/content/tacotron2/wavs", i + ".wav"))

print(listtxt)
with open("/content/tacotron2/filelists/list.txt", "w") as f:
  f.write("\n".join(listtxt))

shutil.rmtree("/content/tacotron2/wavs2")

and the datasets themselves: https://drive.google.com/drive/folders/1OKm39u19bKMuaeR3diKeKVVvsXc1ZNYC

300+ voice clips. not sure if it needs to be trained more or if i should go get more audio clips.

Gene from Bob's Burgers.

yea

im working on all of them xd. started louise but uh.. her voice came out... not so good. im getting more audio for her.

it'll probably sound better with ye good olde Uberduck juice, sounds pretty good already

(just look at this one xD. this one is very incomplete. i am getting more audio clips. but this is so funny to me)

by that i tried, and it sounds more like an ai tried do removal instrumental with similar GAN noises or smth like that