Open-Source Frontier Voice AI
VibeVoice combines 60-minute single-pass ASR, streaming-friendly realtime TTS, and the research arc of long-form multi-speaker voice generation into one sharp public release narrative.
Independent editorial guide to the public VibeVoice release materials, demos, weights, and papers.
Structured transcription, streaming synthesis, and long-form voice research in one visual stack.
Model Family
The current public VibeVoice materials center on long-form recognition, realtime synthesis, and the archival context of the original long-form TTS release.
Accepts up to 60 minutes of audio in one run while preserving speaker continuity, timestamps, and domain-aware phrasing through customized hotwords.
A deployment-friendlier realtime model that supports streaming text input, approximately 300 millisecond first audible latency, and longer running speech sessions.
The original research release introduced up to 90 minutes of generated speech with support for as many as four distinct speakers, but the code was later removed from the repository after a responsible-use review.
Capability Matrix
Instead of presenting voice AI as a single monolith, VibeVoice exposes separate model lanes for long-context ASR, realtime generation, and long-form research synthesis.
The public ASR release emphasizes speaker identity, timestamps, and transcript content as one unified result surface.
VibeVoice-Realtime-0.5B narrows the path to quick experimentation with streaming text input and faster audible response.
The model family is framed publicly as research-oriented, with explicit cautions around production use and synthetic audio misuse.
Release Timeline
The current public story is best understood as a sequence of model releases, demos, and responsible-use adjustments rather than a single product drop.
The public materials note a Hugging Face Transformers release path for VibeVoice-ASR integration.
Single-pass 60-minute transcription, multilingual support, and a public Playground become the clearest onboarding path.
The lighter realtime model adds a public Colab route for quicker hands-on evaluation.
The release history explicitly calls out a change in repository scope tied to safer usage boundaries.
Up to 90 minutes of generated speech with multiple distinct speakers sets the tone for the broader VibeVoice family.
Resources
This site packages the public entry points into one fast overview page, while the underlying model details remain with the official repository and papers.
Browse release notes, docs, and model references from the public Microsoft repository.
Models Hugging Face CollectionMove directly into the public weight pages for ASR, Realtime, and archived TTS artifacts.
Demo ASR PlaygroundUse the clearest public try-now path for the current VibeVoice release stack.
Demo Realtime ColabOpen the streaming TTS notebook to explore low-latency speech generation experiments.
Paper ASR Technique ReportRead the structured transcription and long-context ASR framing behind the public release.
Paper TTS Research ReportReview the original long-form multi-speaker synthesis research that seeded the family.
FAQ
Short answers for what VibeVoice is, what you can try, and how the current scope should be interpreted.
VibeVoice is a family of voice AI models that spans long-form ASR, realtime streaming TTS, and the historical research release of long-form multi-speaker TTS.
The public ASR Playground and the Realtime Colab are the most direct, live entry points surfaced by the current release materials.
No live TTS demo is presented here. The TTS release remains part of the public research story, but the repository notes that the code was removed after a responsible-use review.
The public release guidance presents VibeVoice as research-first and recommends further testing before real-world or commercial deployment.