Microsoft VibeVoice ASR: live speech to text

On 2 September Microsoft Research uploaded VibeVoice-ASR-Streaming to Hugging Face in two sizes, under the MIT licence. The model turns speech into text as it flows, separates the speakers and accepts a custom list of names and terms. There are ten languages, and Bulgarian is not among them.

In short

Two variants, 1.5B and 7B, both under MIT.
Streaming transcription with speaker attribution and custom hotwords.
Languages: Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish.

Checked on1 October 2026Responsible editorTsvetelin IvanovHow we workMethod · Corrections

Meeting minutes that write themselves while people talk, and know who said what.

There are plenty of speech recognition (ASR) models. What is interesting here is the combination: real-time streaming, speaker attribution and an open licence that allows commercial use.

The facts: on 2 September 2026 Microsoft Research published the models VibeVoice-ASR-Streaming-1.5B and VibeVoice-ASR-Streaming-7B on Hugging Face under the MIT licence. According to the team's description, it is a unified streaming speech recognition model that transcribes which speaker says what as speech arrives, and accepts custom hotwords, such as names and technical terms, for better recognition of domain-specific content. It supports ten languages: Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. The code is in the microsoft/VibeVoice repository on GitHub.

Bulgarian is missing. Russian is there. For anyone who has run a foreign speech model on our language, that is no surprise. Open weights do not learn a new language on their own.

Before you read the licence, read the list of languages.

There is a practical upside, though. Custom hotwords do their job exactly where general models stumble: names of people, products, places. If you are building a tool for international meetings in English, there is something to take from here today.

For Bulgarian speech this is not the answer. The licence allows someone to fine-tune it on Bulgarian, and that is the next thing we will be looking for in the repositories.

The visual is generated code art. No third-party images.

Microsoft released an open model that transcribes who says what in real time, but not in Bulgarian

More on this

Cloudflare gives agents a workspace, a small model for decisions and a search that starts costing money in November

Rivals on the same team: 40 companies launch an open alliance for AI agent security

Thinking Machines released Inkling - an open model they don't even market as the strongest