got bored and tried doing a mashup girl voice in elevenlabs. results were... interesting to say the least.

how to learn machine learning?

dammmnnnnnn how da hail you do these way to realistic voicelines?

elevenlabs

requires money now afaik

.

damn just saw that

did some free voices

saw the limits

was like ''HAIILLL NAWWW I WILL SEND YOU TO JESUS''

@TheRoyalRuby2000 So when you made your Wiggles voice models, how much data did you use? I'm giving data collection another shot and last time I generated a Greg voice from 10 minutes of audio with incoherent garbling

Hey man, sorry for the ping but I want to let you know everything. Some of the Wiggles voice models were used with 8 wavs with a little over a minute of entire voice data. But most of them now are used with 4 wavs and they vary depending on how many seconds each wav is. Half are 4 wavs with 10 seconds, with a little under a minute of voice data, and half are 4 wavs with 5 seconds long with close or a little above 30 seconds of entire voice data. I train using Pipeline so that gives an idea. I always use batch size 1. Learning rate depends on how many seconds and how many wavs there are. For the 8 wav datasets, I use learning rate 0.0007. For the 4 wav datasets with 10 seconds for each wav, I use learning rate 0.0004. And for the 4 wav datasets with 5 seconds for each wav, I use learning rate 0.0002.

Can you make sure they all say "G'day mate!" properly?

The problem with that is that I used voice data that doesn't include that phrase so you may need to improvise while synthesizing speech.

fuck

It's not a phrase they say

...maybe it pops up in that Steve Irwin collaboration they did, but I only remember Irwin using the phrase there

I think I've heard Greg use it somewhere

Actually, Jeff said it to Greg in Top of the Tots

At the moment Greg isn't my highest priority I'm spending the rest of February focused on Rias

Oh so I found Tortoise

And uh I managed to make voice clips of Rias speaking in English

So the result is a bit stodgy

I think I need to get at least a few more episodes' worth of Rias audio if I want to get better results

It could also allow me to quickly do Greg, though I imagine there's a better voice cloning software for that It's been awhile since I've looked into these things

Elevenlabs is terrible with non-American sounding voices

weird, since the werehog and suketoudara are from American voice actors. I hope it gets improved.

Did you read it as “Elevenlabs is terrible with American sounding voices”

character voices are also very bad

its likely not gonna be fixed because its not meant for that