VibeVoice

VibeVoice is a novel framework designed for generating expressive, long-form, multi-speaker conversational audio from text. It employs a next-token diffusion framework and an LLM to manage dialogue flow and acoustic details, capable of generating up to 90 minutes of speech.