Phoneme-Level BERT for Enhanced Prosody of Text-to-Speech with Grapheme Predictions

StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models