VibeVoice Technical Report research paper summary (October 2026)

The VibeVoice Technical Report from Microsoft describes a model that generates long, multi-speaker conversational speech. Microsoft has since removed that model's code from the repository, which now offers long-form and streaming speech recognition and a lightweight real-time text-to-speech model.

The report introduces next-token diffusion for speech and a continuous tokenizer that the authors say compresses 80 times more than Encodec at comparable quality. With a 64K context window the model synthesizes up to 90 minutes of speech with up to four speakers.

Since then Microsoft has added VibeVoice-ASR for 60-minute single-pass transcription in over 50 languages, a streaming and a CPU variant of the ASR model and the real-time speech model VibeVoice-Realtime-0.5B. On September 5, 2025, Microsoft removed the VibeVoice-TTS code from the repository after misuse, and VibeVoice-TTS-1.5B is marked as disabled. The repository is under MIT, had 54,529 stars and was updated on September 3, 2026.

• Reads text aloud in real timeVibeVoice-Realtime-0.5B turns streaming text into speech, with first audio after about 300 milliseconds and passages of around 10 minutes, according to the README. Product teams can prototype spoken answers for assistants; the README marks the voices in languages other than English as experimental.
• Transcribes long meetings in one passVibeVoice-ASR processes up to an hour of audio at once and returns who spoke, when and what, with custom hotwords for names and terms. Teams get consistent speaker labels across long recordings.
• Runs speech recognition on CPUsA BitNet-quantized ASR variant runs in real time on CPUs according to the README. IT teams can transcribe on standard servers or edge devices without GPUs.

• One repository covers long-form speech recognition and a lightweight real-time speech model.
• MIT license from Microsoft.
• Microsoft removed the long-form TTS code after misuse, so the model described in the report can no longer be run from the repository.

The paper is freely available on arXiv, and the accompanying code runs on your own hardware in any country.

Microsoft calls VibeVoice a research framework and does not recommend commercial or real-world use without further testing. Synthetic voices can be misused, so set internal rules for consent and labeling of generated audio. Transcripts of meetings contain personal data, so decide where the model runs and how long outputs are kept.

No reviews yet. Be the first to share how your team uses it.

Every review shows its verification level. We read each one before it goes live. How we check reviews