VibeVoice: Turn an Hour of Audio Into a Full Transcript in One Go

VibeVoice is a family of open-source voice AI models from Microsoft, and the one most people will reach for is VibeVoice-ASR, a speech recognition model built for long audio. Feed it up to 60 minutes of continuous audio and it produces a structured transcript in a single pass: who spoke, when they spoke, and what they said. No chopping the recording into 30-second pieces first, no stitching transcripts back together afterwards.

That single-pass design is the whole point. Most speech-to-text tools slice long recordings into short chunks, and every cut is a chance to lose track of who is speaking or what the conversation was about. VibeVoice-ASR accepts the full hour inside a 64K token window, so speaker tracking and context stay consistent from the first minute to the last. It is a Python project, licensed under MIT, and the repository had 55,376 stars and 6,241 forks when we checked on November 5, 2026.

The headline feature is simple to describe and hard to do well: one hour of audio, one pass, one transcript. A 60-minute podcast episode, a team meeting, a lecture, or an interview goes in whole, and what comes out is a structured document rather than a wall of undifferentiated text. Every line carries a speaker label and a timestamp, so you can search for the moment a specific person said a specific thing instead of scrubbing through an hour of audio. An hour of conversation runs to roughly 8,000 words, and pasting the transcript into the Word Counter gives you its length and reading time before you decide how to summarize it.

Under the hood, the VibeVoice family uses continuous speech tokenizers, acoustic and semantic, running at an ultra-low frame rate of 7.5 Hz, which keeps long sequences computationally manageable without throwing away audio fidelity, and the ASR model takes the full hour inside a single 64K token window. The practical result is that long recordings stay coherent: the transcript at minute 55 still knows who everyone is and what they were talking about at minute 5.

A transcript is only useful if you can trust its structure, and VibeVoice-ASR does three jobs at once: speech recognition, speaker diarization, and timestamping. The output answers the three questions that matter for anyone working with recorded speech: Who, When, and What. That combination is what makes it genuinely useful for meeting notes, podcast show notes, research interviews, and lecture summaries, where knowing which person said something is as important as the words themselves.

You can also feed it customized hotwords, names, technical terms, or background context, to guide recognition. If your recording is full of product names, people's names, or domain jargon that generic models always mishear, hotwords give the model a nudge in the right direction and noticeably improve accuracy on the words that matter most to you.

VibeVoice-ASR is natively multilingual, supporting over 50 languages out of the box. That matters because most transcription tools are tuned for English first and treat everything else as an afterthought. Here the multilingual support is a core design feature, not a checkbox, and the project documents its language coverage in the ASR docs.

There is also a streaming variant, VibeVoice-ASR-Streaming, that transcribes while the audio is still arriving, emitting text once per audio chunk instead of waiting for the recording to end. That makes it usable for live captioning and real-time meeting notes, not just post-processing. And if you want to run it on modest hardware, there is VibeVoice-ASR-BitNet, an edge CPU inference engine that compresses the model from 4.62 GB to 1.58 GB and runs in real time on three or more CPU threads, no GPU required.

Transcription is only half the family. VibeVoice-Realtime-0.5B is a lightweight real-time text-to-speech model that streams text in and speaks it out, with around 300 milliseconds to the first audible sound. It is a small 0.5B parameter model, which makes it deployment-friendly, and it handles robust long-form speech generation of around 10 minutes.

The Realtime model also ships with experimental voices, including multilingual voices in nine languages and eleven distinct English style voices, so you can try different tones without retraining anything. You can try it in a Colab notebook straight from the repository, which is the fastest way to hear what it sounds like before you install anything.

The voice side is small for a reason. VibeVoice originally shipped a larger text-to-speech model, and in September 2025 Microsoft removed the VibeVoice-TTS code from the repository after finding it used in ways inconsistent with its stated intent. The README is direct about the risk that remains: convincing synthetic speech can be used for impersonation, fraud and disinformation, so disclose AI-generated audio whenever you share it.

The fastest way in is the ASR Playground linked from the repository, which lets you test transcription without installing anything. If you want to run it yourself, the models are on Hugging Face, including the 7B ASR model, the streaming variant, and the BitNet CPU version, and the ASR model is also available directly through the Hugging Face Transformers library. VibeVoice-ASR is integrated into Azure AI Foundry Labs as well, if you prefer to explore it there.

Developers get more: vLLM inference is supported for faster serving, the finetuning code is open, and there is a full technique report for anyone who wants to understand how the model works. Start with a recording you already have, a podcast episode or a meeting you recorded, run it through the Playground or the model, and check whether the speaker labels survive a full hour. That is the test that separates this from chunk-and-stitch tools.

VibeVoice is a research framework from Microsoft, and the project is upfront about that: it is intended for research and development, and the team does not recommend using it in commercial or real-world applications without further testing. Transcription can still be wrong, biased, or inaccurate, and the model inherits the limits of its base model, so check important transcripts rather than trusting them blindly. If you have the script or notes the speakers worked from, a text diff against the transcript shows exactly which words the model got wrong.

The other thing to know is that the project is actively evolving. New variants land regularly, a streaming model in September 2026, the BitNet CPU engine in July 2026, and the model list in the repository is the current truth, so check the README before assuming a particular model or demo is still there. What is available today is a genuinely strong open-source transcriber for long audio, free under MIT, and at 55,376 stars it is clearly the tool a lot of people reach for when an hour of audio needs to become text.