VibeVoice-TTS

Microsoft's MIT-licensed VibeVoice TTS: a 1.5B model for up to 90 minutes and 4 speakers, plus a 0.5B realtime model. Research use.

In short. VibeVoice turns text into long, natural conversations, such as a podcast with up to four speakers, and a small realtime version starts speaking in about 0.3 seconds. The weights are MIT-licensed, but Microsoft's cards limit the intended use to research, and Microsoft removed the original TTS code from GitHub after seeing misuse.

Summary

VibeVoice is a Microsoft Research text-to-speech framework. Its first model, VibeVoice-1.5B, was open-sourced on 25 August 2025 and generates expressive, long-form, multi-speaker conversation, up to about 90 minutes of audio with up to four distinct speakers, according to its card. It uses continuous speech tokenizers at an ultra-low 7.5 Hz frame rate, a Qwen2.5-1.5B language model for text and dialogue flow, and a diffusion head that produces the acoustic detail, a design called next-token diffusion. It was trained with context lengths up to 65,536 tokens, and the technical report is arXiv 2508.19205. The card lists a larger VibeVoice-Large with a 32K context as disabled.

On 5 September 2025 Microsoft said that it had found uses of the tool inconsistent with its stated intent and had removed the VibeVoice-TTS code from its GitHub repository. The weights remain on Hugging Face under the MIT license, but the model card limits the intended use to research and warns about deepfakes and disinformation. On 3 December 2025 Microsoft released VibeVoice-Realtime-0.5B, a lightweight model with streaming text input that produces its first audible speech in about 300 ms, depending on hardware. It has a single speaker and is intended for English; the card offers nine more languages (German, French, Italian, Japanese, Korean, Dutch, Polish, Portuguese and Spanish) only for exploration.

Because of the research-only intent and the withdrawn code, treat this family as something to study and prototype with, not as a default production choice. For speech generation under a permissive license that is built for products, look at Qwen3-TTS or VoxCPM2, both Apache 2.0.

Good for

  • Research on long-form, multi-speaker speech such as podcast-style dialogue with the 1.5B model
  • Low-latency English narration that starts speaking while an LLM is still streaming its first tokens, with the realtime model
  • Experiments with next-token diffusion speech models whose weights are MIT-licensed

Watch out for

  • The cards limit the intended use to research, and Microsoft removed the TTS code from GitHub in September 2025 after misuse, so check what is still available before building a product
  • The realtime model has one speaker and is intended for English only, and other languages may give unpredictable results
  • Multi-speaker conversation needs the 1.5B model, which is not the streaming one
  • Synthetic speech can be misused for impersonation, and Microsoft's cards warn of deepfake and disinformation risks

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of VibeVoice-TTS. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
VibeVoice-Realtime-0.5B BF16 1.02B 2.04 GB 2.35 GB Laptop or phone
VibeVoice-1.5B BF16 2.7B 5.41 GB 6.22 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.

More about MIT

See also

Frequently asked questions

What is VibeVoice-TTS?
VibeVoice turns text into long, natural conversations, such as a podcast with up to four speakers, and a small realtime version starts speaking in about 0.3 seconds. The weights are MIT-licensed, but Microsoft's cards limit the intended use to research, and Microsoft removed the original TTS code from GitHub after seeing misuse.
Can I use VibeVoice-TTS commercially?
Yes. VibeVoice-TTS is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is VibeVoice-TTS open source?
VibeVoice-TTS is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run VibeVoice-TTS on my own computer?
Yes. The smallest variant, VibeVoice-Realtime-0.5B, needs about 2.35 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was VibeVoice-TTS released?
VibeVoice-TTS was released by Microsoft on August 25, 2025.
How long a context window does VibeVoice-TTS have?
VibeVoice-TTS supports a context window of up to 64K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3-TTS Alibaba Qwen 910M – 1.92B – Apache 2.0 (Permissive) 3.68 GB Laptop or phone
VoxCPM2 OpenBMB 2.29B – Apache 2.0 (Permissive) 8 GB Laptop or phone
OmniVoice Xiaomi 613M – CC BY-NC (Non-commercial) 3.75 GB Laptop or phone
VibeVoice-ASR Microsoft 2.81B – 8.67B 64K MIT (Permissive) 6.47 GB Laptop or phone