ekwek

/

Soprano-80M

like 238

+1

+1

+1

These thieves never release the code

ahahah why thieves? Btw if you are interested in general approach to training here is how style TTS was trained (base for Kokoro TTS): https://colab.research.google.com/drive/1PdcCcTUUYAxBM3_mDT2_UNz0ZoUkMee_?usp=sharing

Obviously, I saw this notebook a million times. But I need their exact training (preferably Soprano, cause it’s fast).

StyleTTS2 uses diffusion. I need without diffusion. Basically, VITS-2 but faster and better! Must be 48 kHz or trainable from scratch on LJSpeech-sized dataset, and phoneme-based only!

And thieves—because they didn’t told where the data came from (not open sourced the full architecture)

Hi everyone, finetuning/training support is one of the most requested features for Soprano, so I will be uploading training code soon!