Whisper made free transcription normal. Microsoft's VibeVoice goes further: it transcribes a full hour of audio in one pass — with speaker labels and timestamps — and it can also talk back to you in real time. It is open source under the MIT licence, the repository has more than 57,000 stars on GitHub, and you can try both sides of it free today, no strong PC needed.
What is VibeVoice?
VibeVoice is a family of open-source voice AI models from Microsoft Research. It has three parts, and it helps to know which one does what:
- VibeVoice-ASR (7B): speech-to-text built for long audio. It takes up to 60 minutes in a single pass and returns a structured transcript with who said it (speaker labels), when (timestamps) and what (the words) — in more than 50 languages, with no language setting needed. You can also give it custom hotwords — names and technical terms — so it spells them right.
- VibeVoice-Realtime (0.5B): a small, fast text-to-speech model that streams speech as the text arrives, with about 300 milliseconds to the first sound and robust generation up to about 10 minutes. It ships with experimental voices in nine languages and 11 distinct English styles.
- VibeVoice-ASR-BitNet: a 1.58 GB version of the transcription model squeezed to run on an ordinary CPU — real-time speed on 3 or more CPU threads, no GPU required.
One honest note first: the original long-form multi-speaker TTS model (VibeVoice-TTS-1.5B, up to 90 minutes with 4 speakers) had its code removed from the repository in September 2025 after Microsoft found it used in ways it did not intend. Its weights are still downloadable from Hugging Face, but the hosted demo is disabled — so this guide uses the two parts that are actually live and supported today: ASR and Realtime.
Method 1: Transcribe free, no install (ASR Playground)
The fastest way in needs no setup at all.
- Open the model page at huggingface.co/microsoft/VibeVoice-ASR and click the Live Playground link.
- Upload your audio — a meeting recording, an interview, a lecture, up to 60 minutes long.
- Add custom hotwords if your audio has unusual names or technical terms.
- Run it and copy out the transcript: every line labelled with its speaker and timestamp.
This is the method to use for podcasts, meetings and interviews. Unlike most free transcribers that chop audio into short chunks and lose track of who is who, VibeVoice reads the whole hour at once, so speaker tracking stays consistent from the first minute to the last.
Method 2: Real-time voice on free Google Colab
Want the talking side? Microsoft ships an official Colab notebook for the Realtime model — it runs on a free Colab GPU.
- Open the notebook: VibeVoice Realtime on Google Colab. Sign in with a Google account.
- In Runtime → Change runtime type, choose a GPU (the free T4 is enough for this 0.5B model).
- Run the cells in order — the notebook installs the libraries and loads the model for you.
- Type or paste your text, pick a voice (try the multilingual voices or one of the 11 English styles), and generate. Speech starts in about 300 ms, so it feels like a live voice, not a file render.
Remember Colab's free tier has daily usage limits and can disconnect idle sessions, so keep your sessions short and download your audio as you go.
Method 3: Run it on your own PC
If you want it local, clone the repository and follow the README's installation steps:
git clone https://github.com/microsoft/VibeVoice
cd VibeVoice
For transcription on a machine with a decent GPU, use the ASR model from the repository. For a machine with no GPU at all, use the BitNet build instead — the CPU engine linked from the README runs the same transcription on 3+ CPU threads. The ASR model is also built into the Hugging Face Transformers library, so Python developers can load it with the standard Transformers speech-recognition tools, and there is a vLLM guide in the repo for serving it at speed.
Honest limits
- Research, not production: Microsoft says VibeVoice is intended for research and development, and does not recommend commercial or real-world use without further testing.
- Transcripts can be wrong: check important transcripts rather than trusting them blindly — especially names, numbers and domain terms. Hotwords help, but they are not magic.
- Use voice responsibly: the 1.5B TTS code was removed for a reason. Convincing synthetic speech can be misused for impersonation and fraud — disclose AI-generated audio whenever you share it.
- The project moves fast: new variants land regularly (streaming ASR in September 2026, the BitNet CPU build in July 2026), so check the README before assuming a particular model or demo is still there.
Bottom line
If you need long audio transcribed free, VibeVoice-ASR is the strongest open option right now: a full hour, speaker labels, timestamps, 50+ languages, no install. And if you want a free real-time voice, the 0.5B model on Colab takes minutes to hear. Open the Playground, upload one recording, and you will know within minutes whether it earns a place in your workflow.
Links: VibeVoice on GitHub · VibeVoice-ASR on Hugging Face · Realtime TTS Colab notebook · VibeVoice-ASR technical report (arXiv:2601.18184). Hero photo: close-up of a microphone in a recording studio, by Shixart1985, CC BY 2.0, via Wikimedia Commons.
Originally published on Ai Game Changer Pro.
Top comments (0)