Microsoft Open-Sources VibeVoice: 90-Minute Multi-Speaker Speech Synthesis, ICLR 2026 Oral

Microsoft Research / HuggingFace·medium signal

Microsoft Research open-sourced VibeVoice, a next-token diffusion model for long-form speech that synthesizes up to 90 minutes with 4 speakers in a 64K context window. Key innovation: continuous speech tokenizers at 7.5 Hz (ultra-low frame rate) preserving fidelity while enabling long sequences. Handles breaths, lip smacks, and natural turn-taking. Available on HuggingFace as VibeVoice-1.5B and VibeVoice-Realtime-0.5B. Accepted as ICLR 2026 oral presentation.