it makes these noises for some reason

turn off the pronounctation thing

Thanks

84 files were used here, I should have collected more

is that gabe newell

yes

In general, most of the files were taken from dota 2

do you think the models ready or not?

mmmmm

the graph is still smooth though

thanks for the info. i saw SV2TTS but is there a Google colab with a guide on it? it's not really unseen voices or multiple speakers per se. it's the same actress but two different series i'm sampling from, and her character is different in each series. not radically, but different. 60 files in one and 20 in the other, but the 20 are mostly borderline quality where i had to wash out BG music with RX.

No collab as far as I'm aware, but theres tutorials on how to train locally

Itd work well for that approach, you can get it to generate text in each character based on the input verification sound file

yeah no dice for me on me Mac/hackintosh then - no Nvidia GPU (only AMD), and i did try installing the local version requirements but it failed on a component for pytorch (numba). there is a colab i used based on CorentinJ's work but it seems to only allow one file (it'll open several files, actually but it doesn't seem to use them) interface is definitely different than the local version as well. that's the version i was running on Google Colab with the tantalizing but wildly inconsistent results.

You'll have to modify the collab then

You could also try using Amd's hardware acceleration thing and replace parts of the code

yeah i did see see that pytorch had beta support for AMD but not on macOS. i think i'd rather do it on the colab. i've done scripting but not in python and this is my first foray into TTS really. i was mentioning that it only seemed to use one file based on the output. but for sure it did not seem to have the same amount of configurable options as the local one. code output listed all the uploaded clips but only autoplayed the last one. that may not mean anything. but the synthesizer output seemed like it was only pulling data from one file. i might try again and leave that file out just to see if my hunch is right. but i'm definitely not sure how to deal with all the synthesized misfires (breathing, high pitches, etc) in terms of tweaking the code.

it liked short-ish full sentences 5-12 seconds. i uploaded a few minutes of talking quotes and that was a total dud.

this is the colab i used: https://colab.research.google.com/github/tugstugi/dl-colab-notebooks/blob/master/notebooks/RealTimeVoiceCloning.ipynb#scrollTo=WZjKkvGF1Y-i

i'll try another one i found in a bit.

You could always use a virtual machine

SV2TTS isnt gonna sound good without training

yeah i figured that it needed training. you mean to run it in VirtualBox or something? i do have a Win laptop with a 1070 i could utilize if i needed to. that seems easier than say a Linux VM running python stack with pytorch and the ROCm 4.2 beta (and then having to tweak the code) on the Mac. just seeing if there was a colab method to try first.

You're gonna have to tweak the code regardless, that's why I recommend either making your own collab or running it locally (easy to keep track of changes via git)

well the local install is definitely more involved. in terms of tweaking the code, assuming i get the local version up that means tweaking the python code in the toolbox itself.

the notebooks are adapted (sometimes copied) from the pony preservation project, though i believe the models themselves are just stock tacotron2/talknet

if it was anything beyond "press play and follow the instructions" we likely wouldn't have so many models

i believe the latter yeah

plus the notebook has integrated vocoder training, whereas with tacotron uberduck only added the option recently