shijincai/VibeVoice

Combines continuous speech tokenization at 7.5 Hz with a next-token diffusion framework that leverages LLM understanding of dialogue flow and a diffusion head for acoustic synthesis, enabling synthesis of up to 90 minutes of multi-speaker (up to 4 distinct speakers) conversational audio. The architecture preserves audio fidelity while maintaining computational efficiency for long-form generation, outperforming traditional TTS approaches limited to 1-2 speakers. Supports inference via Gradio demo, direct file-based synthesis, and integrates with HuggingFace and Modelscope for model distribution.

No commits in the last 6 months.