VibeVoice: Microsoft's Free Open-Source Voice AI That Transcribes an Hour of Audio in One Pass

If you work with audio — podcasts, meetings, interviews, lectures, voice notes — you have probably hit the same wall: transcription is either expensive, capped at a few minutes per file, or locked behind an API key and a billing page. VibeVoice, a free open-source voice AI from Microsoft, takes a different approach. It lives on GitHub, it is MIT-licensed, and it is built to handle long audio in a single pass.

It is currently trending on GitHub with around 55,000 stars, and it is worth a look if you build anything with voice. Here is what it does, who it is for, and how to try it.

Get the code

VibeVoice on GitHub

That short link takes you to the official repository at github.com/microsoft/VibeVoice, where the README is the source of truth for setup and the latest status.

What VibeVoice does

Transcribe up to 60 minutes of audio in one pass. This is the headline feature. Instead of chopping a long recording into small pieces and stitching the results back together, you can run an hour-long file through it in a single pass. That matters for real recordings: a full podcast episode, a team meeting, a lecture, or a long interview.

Speaker labels and timestamps. The transcript tells you who said what, and when they said it. That turns a wall of text into something you can actually search, skim, and quote. If you have ever tried to find one moment in a 45-minute meeting recording, you know why this matters.

Support for 50+ languages. The transcription is not English-only. If you work with multilingual audio — or your users do — that is a practical difference, not a marketing line.

Real-time text-to-speech. VibeVoice also goes in the other direction: it can turn text into speech in real time. So the same project covers both sides of voice work — getting text out of audio, and getting audio out of text.

Free and open source, MIT licence. There is no paywall for the core tool. You can read the code, run it, and adapt it, under the terms of the MIT licence.

Who it is for

• Developers building voice features. Transcription pipelines, meeting-note tools, podcast show-note generators, searchable audio archives, and accessibility tooling are all natural fits.
• Content creators and podcasters. A transcript with speaker labels and timestamps is the raw material for show notes, quotes, chapters, and searchable episode pages.
• Students and researchers. Lectures and interviews become text you can search instead of audio you have to replay.
• Multilingual teams. If your audio spans several languages, a tool that handles 50+ of them saves you from juggling a different service per language.

How to try it

Start at the repository linked above and read the README first. It has the setup instructions, the requirements, and examples of how to run the tool. As an open-source project, the README is where the current, correct instructions live — follow it rather than any second-hand guide, including this article.

A sensible first test is a recording you already have: a meeting, a podcast episode, or an interview. Run it through, then read the transcript with the speaker labels and timestamps and judge for yourself whether the output is good enough for your use case. That one test will tell you more than any feature list.

Honest limits

No tool is right for every job, and this one is no exception:

• Check the output. Like any transcription tool, the results depend on the audio you feed it. Noisy recordings, heavy accents, and people talking over each other will make any transcript less reliable. Treat the first run as a draft, not a finished document.
• No benchmarks here. I have not run controlled accuracy or speed tests, and this article makes no claims about how VibeVoice compares with other tools on those measures. The features above are what the project describes; your own test on your own audio is the way to judge it.
• It is an open-source project, not a managed service. There is no support desk or uptime promise. Setup, maintenance, and updates are on you — that is the trade-off for a free, MIT-licensed tool you can run yourself.
• Check the README before you build on it. Requirements and behavior can change as the project develops. If you are planning to depend on it in production, read the repo first and keep an eye on its releases.

Bottom line

Most voice tools ask you to choose: free but limited, or capable but paid. VibeVoice is interesting because it does not force that choice. An hour of audio in one pass, with speaker labels, timestamps, 50+ languages, and real-time text-to-speech — free, open source, and MIT-licensed — is a combination worth trying if you work with voice at all.

Head to the repo, run it on a recording you care about, and see what comes back. If the transcript is good enough for your workflow, you have just found a tool that costs nothing and belongs to no vendor.