xvasynth

Progress update; Another v2.0 voice

Added 2021-10-15 16:28:52 +0000 UTC

Just wanted to make a quick post to update people with what's going on at the moment.

I've been hard at work optimizing the training scripts, and I think I've now got it as good as it's going to get for now. It's definitely faster, now, which is great, but still slower than the v1.0 training script, simply due to the fact that the model is bigger (more components to train), and that there is an alphabet about twice as big to learn (normal + the ARPAbet symbols).

I've gone through a few variations of it, which HAS meant that I've had to re-do the models I had so far, but now that the training pipeline is mostly there, it's now just time to run the training! (currently underway)

The changes I've made have also resulted in actually better quality output too (especially with speech-to-speech), which is definitely an important bonus! And most of this work actually ticks off a lot of the upcoming work I had to do for xVATrainer, so it's definitely been worth it.

I've attached the first final v2.0 model, Mass Effect: Joker.
I'll have an updated/better version of the Skyrim: FemaleEvenToned voice ready soon, too.

I've already put together most of the video for the v2.0 release showcase, so we're very close now to the full release, which will include several tweaks/fixes over the pre-release - thank you for the feedback and bug reports! They're super useful and appreciated!