Training Audio With Tacotron 2 Colab Page
Last updated
Was this helpful?
Last updated
Was this helpful?
Training is the main part of tacotron. If you won't train your voice model, then you can't even load wavs with synthesis.
Now you've may be asking, what is google colab?
You never probably heard of this, but here, I'll try to explain as I can.
Google colab is a VM server (and no, it's not for cryptocurrency mining stuff) that you could train some ai models without even having an expensive GPU. ISN'T THAT WACKY JOSUKE? but here's a thing, it has limits.
Number one, GPUs can be limited. You may encounter some low video memory so, be careful what you training for, or else you'll end up getting out of memory error. Try lowering value things.
Number two, session times can be limited. What I mean by that is that, if you let your colab runtime running up to 12 hours, you will be kicked and you won't be returning until in 2 days.
Number three, colab can be buggy as hell. For example, you upload so many files you have done gathering dialogs. But here's a twist, Colab can deny uploading files after you have waited for 30 seconds. It sucks I know, but it should be working with another browser.
Google colab also has a pro subscription, it's optional but 10 bucks will be that worth (heh, it's the same cost as normal Discord Nitro) and probably will be better. Which means more session times and good GPUs.
Update: They apparently have changed things since now free plan gives you only k80's (and sometimes P100 and V100) So if you can't be that patient i suggest buying pro or just keep clicking on reset factory until you get a faster GPU, or possibly train it on your end if you have a decent GPU (like RTX 30/20 series with somewhere more GB of VRAM) (I can't tell you how right now but someone in discord server might help you with it.)
If you haven't done gather dialogs and transcribing audios then what are you doing? go back where you came here from:
Here are some guides that you might need to read:
If you came across errors then go to this page:
