Microsoft VibeVoice: The Open-Source Frontier of Voice AI

Microsoft's VibeVoice is an open-source voice AI family that combines TTS and ASR. Its core is a 7.5Hz continuous speech tokenizer, along with a next-token diffusion architecture where an LLM understands context and conversational flow while a diffusion head generates acoustic detail.

Recent releases are as follows.

• 2026-03-06: VibeVoice-ASR was included in a Hugging Face Transformers release.
• 2026-01-21: Released an ASR that processes up to 60 minutes of audio in a single pass. It produces structured transcripts capturing speaker (Who), timestamps (When), and content (What) together, and supports customized hotwords.
• Supports 50+ languages, and also provides fine-tuning code and vLLM inference documentation.
• 2025-12-03: Open-sourced VibeVoice-Realtime-0.5B, which supports streaming text input and long utterances.
• 2025-12-16: Added experimental speaker presets, providing 9 languages including German, French, Italian, Japanese, and Korean, along with 11 English-style voices.
• 2025-09-05: Removed the VibeVoice-TTS code from the repository due to responsible-use concerns.
• 2025-08-25: Released VibeVoice-TTS, which synthesizes conversational speech up to 90 minutes long with up to 4 speakers, and it was accepted as an ICLR 2026 Oral.

The model lineup includes VibeVoice-ASR-7B, VibeVoice-TTS-1.5B, and VibeVoice-Realtime-0.5B, along with Hugging Face weights, a Playground, Colab demos, and fine-tuning resources.