Microsoft released VibeVoice, an open-source voice AI family, with 56K stars

A team that pays per minute for a text-to-speech or speech-to-text service now has an open option from Microsoft. Microsoft released VibeVoice on GitHub under the MIT licence. The repository has 56,327 stars and was ranked 7th on Trendshift's daily list on 6 November 2026. Microsoft's own tag line for the project is "Open-Source Frontier Voice AI".

The project is a family of models, not one. The README lists five models under the VibeVoice name. Three are for speech recognition. Two are for text-to-speech. One of the recognition models is a BitNet build meant to run on CPU. One text-to-speech model is a real-time 0.5 billion parameter version. The models are available on Hugging Face and on Azure AI Foundry Labs.

What the models cover

The speech recognition side supports 50 or more languages. The README says a single pass can read up to 60 minutes of audio. It identifies speakers with timestamps. Users can set hotwords that the model should listen for. The text-to-speech side generates up to 90 minutes of speech. It supports multi-speaker voices and expressive output.

The project uses a continuous speech tokenizer at 7.5 Hz. It uses a next-token diffusion framework with a language model for context. The README describes the design as "diffusion plus LLM", which is now a common shape for voice work. Microsoft is the stated builder.

Who it is for

The target user is a team already paying for voice services. Three cases are in scope. First, a product that reads text aloud, like an accessibility layer or a reading app. Second, a product that transcribes voice, like a meeting tool or a call-centre app. Third, a product that holds a real-time voice conversation with an agent.

The licence matters here. The MIT licence means a team can use the models in a product it sells. It can run them on its own machines or on a cloud it already pays for. The open release is the point of this launch. Microsoft was not the first to release open voice models. It is the first large US cloud to release a full family at this size, built for both tasks.

What is not yet clear

The README does not state a benchmark score against a paid service like ElevenLabs, OpenAI Whisper, Deepgram or AssemblyAI. The README does not give a cost per minute for running the model on a given GPU. A team that wants to move off a paid service has to run the models on its own hardware. The team has to measure quality and cost on its own voice data.

The honest test for VibeVoice is whether the output stands up on a team's own audio. A demo on English podcasts says nothing about a sales call in strong accents, or a reading app for children. A team should pick five minutes of its own hardest audio. Run it through VibeVoice and through the current paid service. Compare the output. The decision takes an hour and does not need a model licence.

Source

• Primary source: microsoft/VibeVoice on GitHub.