Spaces:

---

SupraLabs

/

README

Like 3

Running

Hey !
I've been following your work for a while now. And I recently saw that you made a TTS model.
And I've been making SOTA TTS models for a few months now.

I’d love to collaborate with SupraLabs on efficient TTS. My model, Vocetta2-1m (https://huggingface.co/VantoraLabs/Vocetta2-1m), achieves 4.04 SCOREQ, 3.97 UTMOS and 3.64 DNS-SIG with just 1.03M parameters (better than TinyTTS, Inflect Nano, Kitten TTS nano, SupraTTS-0.1-Beta and relatively close to models dozens of times its size) and runs at 97× real-time on my Ryzen 5 3500x (faster than ANY neural TTS model I know of).
It uses a three-stage, teacher-distilled architecture, with most parameters dedicated to its waveform decoder. Key techniques include checkpoint averaging and multi-metric evaluation.

The MIT-licensed code is available, and I’d be happy to share training recipes, benchmarks, and expertise to help advance SupraLabs’ TTS research.