Ah fair play. Sounds good.

GlowTTS is also slightly promising for fast multispeaker synthesis

I feel like the architecture I'm looking for is still in development and I'm not adept enough at ML atm to construct it myself

TTs is kinda the most fun part, for me atleast

Yea, its a tricky one. Considering not all of the papers code is made public, requiring others to approximate and share the code. And yea lots of the people who really know what their doing will be those behind the dozens of new tts startups lol

If I wanted to go super high effort, I could generate character speech using DialoGPT trained on character dialogue, hook it up to a finetune TTS architecture, and use STT to allow you to talk to the character over voice

But I'll probably just combine a transformer + TTS

Unless I go for a small quick model itd be pretty slow on my hardware

I mean for me a colab has to be there for me to try to test out an interesting repo 😅 . Its terrible that I just dont know how to figure out how to run everything myself or just make my own colab lol.

Try learning how to setup things locally

The collabs make sense if you're training cuz your gpu will be way worse at that

But for inference or using it on other things after its trained you can get it decently fast using fp16 optimization + cuda integration

I'm on a cruddy laptop so can't really run things locally unfortunately lol. I really just need to learn how to implement code into colab.

Hmmm

How much vram does your laptop's gpu have?

Have you seen this? It's cool af and seems pretty easy to implement. You could then just use prexisting audio for style tokens? Even if its the same data I think it makes it sound more natural doing it that way

Got ~6gb of vram and 6gb of free Ram

My gpu had 6gb vram too

You can run small models on it

Running on an SSD so it aint too bad a setup. Im just also very low on storage which doesn't help.

SV2TTS can even train on 6gb Vram if you decrease the batch size by 1/2

So yea that's good enough

Hmm ok

Yea thats interesting

Vram is often the bottleneck w/ training and especially transformers

All the SV2TTS samples Ive heard though do sound a lot worse than any of the Tacotron2 models Ive made

Running your models with cuda support speeds it up a ton

Have you seen the one I posted in #795444484242145302 ? How does that compare?

The attention on it seems a big scuffed tbh

Unfortunately bought my laptop before I knew what Cuda was :/. So just got a cruddy AMD Ryzen