Li et al. · 2023 · style diffusion
Styles sampled with diffusion, then a decoder, with WavLM as an adversarial ear. Reported human-level MOS on LJSpeech. Kokoro’s architecture is a skinny descendant. The clip is the authors’ LJSpeech demo — same sentence as the VITS and NaturalSpeech clips from that page.
Human-level MOS on a single-speaker corpus, and the recipe Kokoro slimmed down.
Vendor / paper demo
LJSpeech (in-distribution) · StyleTTS 2 official demo page, LJSpeech ID 00005. Same sentence as VITS and NaturalSpeech on that page.
After the construction and action of the machine had been explained, the doctor asked the governor what kind of men he had commanded at Goree,