VALL-E is a neural codec language model using discrete codes derived from an off-the-shelf neural audio codec model, and regard TTS as a conditional language modeling task rather. VALL-E emerges in-context learning capabilities and can be used to synthesize high-quality personalized speech with onl…