Microsoft releases VibeVoice-Realtime-0.5B, a streaming TTS model under MIT
A single-speaker streaming TTS model that starts speaking in about 300 ms; MIT licensed, but the card says research use only.
Microsoft published VibeVoice-Realtime-0.5B, a text-to-speech model built on a Qwen2.5-0.5B language model. The card says it takes streaming text input and produces first audible speech in about 300 ms, depending on hardware. It uses an acoustic tokenizer at 7.5 Hz with no semantic tokenizer, and a diffusion head conditioned on the language model's hidden states. Context is 8k tokens, with generation of about 10 minutes.
The card reports a word error rate of 2.00 and speaker similarity of 0.695 on LibriSpeech test-clean, and 2.05 and 0.633 on SEED test-en. It supports a single speaker. It is primarily English; nine other languages are offered "for exploration" and the card also says it is intended for English only.
The licence is MIT. The card also says the model is "intended for research and development purposes only" and advises against commercial or real-world use without further testing. Out-of-scope uses include voice impersonation without recorded consent. The card lists 0.5B parameters while the Hub metadata shows 1B, and does not explain the gap.
The card gives no release date. The date used is the Hugging Face repo creation date.
An MIT licence with a research-only warning is a mixed message. Read both.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights are public under MIT; the card's own research-only wording limits real use.
- Useful 3: One English speaker and no cloning keeps it narrow.
- Notable 3.5: Over 1,300 likes on Hugging Face and a big-vendor name; the card's own limits temper it.
- Novel 3: Streaming text input with a 7.5 Hz acoustic tokenizer and diffusion head is a distinct design.
Comments