Development timeline

From the first TTS release to streaming, speaker-attributed recognition — and a build that runs on a CPU.

  1. 25 Aug 2025

    TTS open-sourced

    VibeVoice-TTS debuts: up to 90 minutes and 4 distinct speakers in one pass. The work is later accepted as an ICLR 2026 Oral.

  2. 5 Sep 2025

    TTS code removed

    After finding uses inconsistent with the stated intent, Microsoft pulls the TTS code from the repository. The weights remain on Hugging Face.

  3. 3 Dec 2025

    Realtime-0.5B open-sourced

    A real-time text-to-speech model that takes streaming text input and still holds together over long passages.

  4. 16 Dec 2025

    More Realtime voices

    An experimental speaker pack adds nine multilingual voices (DE, FR, IT, JP, KR, NL, PL, PT, ES) and 11 English style voices.

  5. 21 Jan 2026

    ASR open-sourced

    A unified speech-to-text model: 60 minutes in a single pass, structured into who, when and what, across 50+ languages. Fine-tuning code and vLLM inference follow.

  6. 6 Mar 2026

    ASR in Transformers

    VibeVoice-ASR ships as part of a Hugging Face Transformers release, so it can be called like any other model.

  7. 12 Mar 2026

    ASR in Azure AI Foundry Labs

    The model becomes explorable and testable through Microsoft Foundry, without a local setup.

  8. 23 Jul 2026

    ASR-BitNet — off the GPU

    Heterogeneous quantization compresses ASR from 4.62 GB to 1.58 GB, hitting a real-time factor under 1 on three CPU threads. Shipped with the VibeASR.cpp inference engine.

  9. 3 Sep 2026

    ASR-Streaming released

    The newest member: 1.5B and 7B models that transcribe who said what while the audio is still arriving, with hotwords and 10 languages.

中文版