LATEST
AI Tools

Microsoft VibeVoice Review: Free Voice AI, and the Cloner It Pulled

Updated November 17, 2026
9 min read
Editor's Note: This article was human-reviewed for accuracy
Close-up of a professional studio microphone with a pop filter in a soundproof studio

Quick Answer & Key Takeaways

  • • Microsoft's VibeVoice is a free, open-source voice AI family that turns hour-long meetings into speaker-labelled transcripts in a single pass.
  • • It runs on your own hardware, so there is no per-minute bill and no audio leaving your machine.
  • • The catch is the voice-cloning half: Microsoft removed the TTS code in September 2025 after misuse, so the headline feature no longer ships in the repo.

Microsoft's VibeVoice has more than 57,000 GitHub stars, costs nothing, and promises something no hosted transcription service offers: hour-long meetings turned into speaker-labelled transcripts on your own machine, with no audio leaving it. The transcription half of that pitch is real. The other half, the voice cloning that made VibeVoice famous in August 2025, no longer ships in the repository. Microsoft pulled the text-to-speech code a month after release, after discovering uses it calls inconsistent with the project's intent. What follows is what VibeVoice is in November 2026, what still works, what it costs to run, and whether Microsoft's own accuracy numbers hold up.

What VibeVoice actually is

VibeVoice is a family of open-source voice models from Microsoft Research, published under the MIT licence in the microsoft/VibeVoice repository. It started in August 2025 as a text-to-speech system and has since grown into five pieces: VibeVoice-ASR for transcription, ASR-Streaming for live transcription, ASR-BitNet for CPU-only transcription, VibeVoice-TTS for long-form speech synthesis, and Realtime-0.5B for streaming speech synthesis.

VibeVoice uses continuous speech tokenizers running at 7.5 Hz, which is what lets one model hold an hour of audio in context.

It had 57,218 stars and 6,464 forks when checked on November 17, 2026.

VibeVoice-ASR: the part that works today

VibeVoice-ASR, released on January 21, 2026, is the model most people should care about. It is a 7-billion-parameter speech-to-text model that processes up to 60 minutes of audio in a single pass, inside a 64,000-token window, and returns a structured transcript with three things most transcription tools handle separately: who spoke, when they spoke, and what they said.

Three details make it different. First, it supports more than 50 languages with no language setting required, and it handles code-switching, where a speaker moves between languages mid-sentence. Second, it accepts customised hotwords, so you can feed it names, product terms, or jargon before transcription. Third, the single-pass design keeps speaker identity consistent across the whole hour.

The streaming variant, released September 3, 2026, transcribes while audio is still arriving and covers 10 languages, in 1.5-billion and 7-billion-parameter versions. Microsoft reports an expected speaker-attribution delay of 2.00 seconds in its 22-frame configuration. The CPU variant, VibeVoice-ASR-BitNet, released July 23, 2026, compresses the model from 4.62 GB to 1.58 GB and runs faster than real time on three or more CPU threads, with no GPU at all. Microsoft's model card puts its real-time factor at 0.42 on eight threads of an AMD EPYC 7V13 server processor.

There is also a no-install way to try it. Microsoft runs a public playground for the transcription model, and the model has been part of Azure AI Foundry Labs and the Hugging Face Transformers library since March 2026. LoRA fine-tuning code is published, and vLLM inference is supported for serving at scale.

The TTS story: what Microsoft pulled, and what is left

The original VibeVoice-TTS, released August 25, 2025, could synthesise up to 90 minutes of conversational speech in one pass, with up to four distinct speakers and natural turn-taking. It was accepted as an oral presentation at ICLR 2026. Then, on September 5, 2025, Microsoft removed the TTS code from the repository, writing that it had discovered instances where the tool was used in ways inconsistent with the stated intent.

What remains of the speech-synthesis side is thinner. The VibeVoice-TTS-1.5B weights are still hosted on Hugging Face, but the repository's quick-try link is disabled and the code to run it is gone. The live piece is VibeVoice-Realtime-0.5B, released December 3, 2025, a half-billion-parameter streaming model with about 300 milliseconds to first audible audio and support for roughly 10-minute generations. Experimental speakers added on December 16, 2025 cover nine languages and 11 English voice styles.

If you want a voice cloner, VibeVoice is no longer the tool for that job, and any guide that still presents it as one is out of date. Our own VoiceStudio review covers a local voice-cloning tool that still ships its full code. VibeVoice today is a transcription project with a small streaming-speech side, not a cloning project.

How to run VibeVoice-ASR on your own machine

Running the full 7-billion-parameter model needs a GPU. Microsoft's documentation recommends starting from NVIDIA's PyTorch container, installing flash-attention if it is not already present, then cloning the repository and installing it:

sudo docker run --privileged --net=host --ipc=host --ulimit memlock=-1:-1 --ulimit stack=-1:-1 --gpus all --rm -it nvcr.io/nvidia/pytorch:25.12-py3
git clone https://github.com/microsoft/VibeVoice.git
cd VibeVoice
pip install -e .

From there, the repository ships two entry points: a Gradio demo with a web interface for uploading audio, and a file-inference script that transcribes audio from the command line:

python demo/vibevoice_asr_gradio_demo.py --model_path microsoft/VibeVoice-ASR --share
python demo/vibevoice_asr_inference_from_file.py --model_path microsoft/VibeVoice-ASR --audio_files [add an audio path here]

The ffmpeg package is needed for the demo. If you have no GPU, the BitNet build is the realistic route: a 1.58 GB model that runs in real time on an ordinary CPU. That is the version to reach for on a laptop, and it is also the version where the accuracy trade-off is least documented.

What it costs

VibeVoice costs nothing to download. The code and model weights are MIT-licensed, with no subscription, API key, or per-minute charge. The real costs are the ones every self-hosted model carries: a GPU with enough memory for the 7-billion-parameter model, or a cloud GPU rental if you do not own one, plus electricity and setup time. For the CPU build, the hardware cost is close to zero on a modern laptop.

That zero marginal cost is the whole pricing story. The table below compares the options using prices from our own recent reviews, with sterling and Canadian dollar equivalents at November 17, 2026 market rates (1 USD = 0.758 GBP, 1.43 CAD).

ToolWhat you payBest forHosted API
VibeVoice-ASR$0, self-hostedLong meetings, private audioNo, run it yourself
ElevenLabs Scribe v3$0.20 per hour (about £0.15 / C$0.29)Hosted transcription, no setupYes
Mistral Voxtral 3$0.04 per minute, $2.40 per hour (about £1.82 / C$3.43)Live phone calls, voice agentsYes
OpenAI Whisper large-v3-turbo$0, open weightsShort clips, quick local jobsNo, run it yourself

The arithmetic favours self-hosting only at volume. At Scribe's $0.20 an hour, transcribing 100 hours of meetings a month costs $20, less than an hour of GPU rental on most cloud providers. VibeVoice wins when the audio is sensitive enough that it cannot leave your machine, or when the volume is high enough that a fixed GPU beats a per-minute meter. It also wins on privacy: the audio never goes to a vendor's servers.

How accurate is it? Read Microsoft's numbers carefully

Every accuracy figure for VibeVoice-ASR comes from Microsoft itself. The model is still absent from Artificial Analysis's speech-to-text board, which had 61 models evaluated as of October 26, 2026, so there is no independent leaderboard number to check it against.

On the multilingual MLC-Challenge meeting set, Microsoft reports an average diarization error rate of 3.42% and a word error rate of 12.07% across 11 languages:

LanguageWord error rateDiarization error rate
English7.99%4.28%
Spanish8.04%2.67%
Korean9.65%4.52%
French15.21%3.80%
German16.30%1.04%
Portuguese21.54%7.98%

English and Spanish meetings transcribe well. Portuguese, at 21.54% word error rate, does not, and French and German sit in the middle. The far-field test, a single distant microphone in a meeting room, is the weak spot: 24.65% word error rate on the AMI-SDM set. That is the scenario most real meeting recordings resemble, a laptop microphone at the end of a table, and it is where the model loses roughly one word in four.

The streaming model compares better against its rivals. In Microsoft's streaming technical report, the 7-billion-parameter model achieved the best or tied-best speaker-attributed error on 12 of 13 evaluation settings, and its five-set mean error of 24.66 beat Gemini 3.5 Transcribe Live at 25.23, GPT Realtime Whisper at 39.31, and ElevenLabs Scribe v2 Realtime at 41.39. Those are Microsoft's own comparisons, run on meeting benchmarks rather than the read-speech sets most leaderboards use, so treat them as a signal, not a verdict. One independent review, checked October 26, 2026, rates the transcription model 3.8 out of 5 and reaches the same conclusion.

Five things to know before you build on it

First, the licence and the README disagree in tone. The MIT licence permits commercial use, but the README says the models are intended for research and development only and advises against commercial use without further testing. Treat the commercial question as unresolved, not settled by the licence alone.

Second, the 60-minute limit is a hard one. Recordings longer than an hour must be split, and the speaker-consistency benefit ends at the split.

Third, the streaming model covers only 10 languages, against more than 50 for the full model. If your live use case is outside those ten, the live option does not apply.

Fourth, the BitNet CPU build's accuracy against the full model is not published. Microsoft publishes its speed, a real-time factor of 0.42 on eight server threads, but not how much accuracy the compression costs. Independent laptop testing found faster-whisper's small models running faster on the same hardware, which suggests the CPU build's advantage is the speaker labelling, not raw speed.

Fifth, do not build a voice-cloning product on this repository. The TTS code was removed for misuse, and the README warns that high-quality synthetic speech can be used for impersonation and fraud. It also asks users to disclose AI-generated content.

Is Microsoft VibeVoice really free?

Yes. The code and model weights are MIT-licensed, with no subscription, API key, or per-minute charge. Your costs are the hardware to run it, electricity, and setup time.

Can VibeVoice clone my voice?

Not from this repository any more. Microsoft removed the VibeVoice-TTS code on September 5, 2025, after discovering uses inconsistent with the project's intent. The TTS weights remain on Hugging Face, but the code to run them is gone.

Can I run VibeVoice without a GPU?

Yes, using the BitNet build. It compresses the transcription model from 4.62 GB to 1.58 GB and runs faster than real time on three or more CPU threads. Microsoft has not published how much accuracy the compression costs compared with the full GPU model.

How accurate is VibeVoice-ASR?

On Microsoft's own MLC-Challenge meeting benchmarks, it averages a 12.07% word error rate across 11 languages, from 7.99% in English to 21.54% in Portuguese. Far-field audio is weaker, at 24.65%. All figures are company-reported.

Can I use VibeVoice in a commercial product?

The MIT licence permits commercial use, but the README says the models are intended for research and development only and advises against commercial use without further testing. Treat that as an open question for your own legal review.

VibeVoice is a strong, free transcription tool with a complicated past. It turns an hour of multi-speaker audio into a structured, speaker-labelled transcript on hardware you own, a job hosted services charge for. The voice cloning that made it famous is gone, and the accuracy numbers, while detailed, are all Microsoft's own. Try the public playground first, and if the English meeting numbers hold up on your audio, the CPU build is the cheapest serious transcription setup you can run today. Our ElevenLabs Scribe v3 review and Retell Tell-1 review cover the hosted alternatives if you would rather pay per minute than run a model yourself.

Sources

9 min read
205 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more