← Back to Discover

VibeVoice

Microsoft's open-source voice models: long-form speech recognition that labels who spoke and when, plus a small real-time text-to-speech model.

Open modelsTranscriptionText to speech
Visit website ↗

What it does

VibeVoice is described on its repository as a family of open-source voice AI models covering both speech recognition (ASR) and text-to-speech (TTS). It uses continuous speech tokenizers running at a very low frame rate, with a language model handling context and a diffusion head generating acoustic detail. The speech recognition side is the main line today. VibeVoice-ASR takes up to 60 minutes of audio in a single pass and returns a structured transcript of who spoke, when and what was said. You can supply hotwords such as names or technical terms to guide recognition, and the site lists support for over 50 languages. A streaming variant transcribes while audio is still arriving, and a BitNet build is meant to run on CPUs without a GPU. The page also lists finetuning code and vLLM inference support. On the text-to-speech side, Microsoft removed the VibeVoice-TTS code in September 2025 after finding uses it called inconsistent with the stated intent, and the model table marks the 1.5B TTS model as disabled. VibeVoice-Realtime-0.5B, a lightweight streaming text-to-speech model with roughly 300 ms to first audio, remains listed. The repository is MIT licensed, but the project says the models are intended for research and development and are not recommended for commercial or real-world use without further testing. It also warns about deepfake and disinformation risks.

Capabilities

  • Speech recognition for up to 60 minutes of audio in one pass
  • Transcripts that mark who spoke, when and what was said
  • Custom hotwords to improve accuracy on names and technical terms
  • Streaming recognition variant and a BitNet variant for CPU inference
  • VibeVoice-Realtime-0.5B for streaming text-to-speech
  • Finetuning code and vLLM inference support for the recognition model

When to use it

  • Research or prototyping with long recordings that need speaker labels and timestamps
  • Running speech recognition on a machine without a GPU
  • Experimenting with low-latency streaming text-to-speech
  • Development work where the project's research-only guidance is acceptable