Microsoft Releases VibeVoice-ASR-Streaming-7B, an Open-Weight Streaming Speech Recognition Model with Speaker Attributio

Microsoft Research has released VibeVoice-ASR-Streaming-7B, an open-weight streaming automatic speech recognition (ASR) model designed to transcribe both speaker identity and spoken content in real time. The model is available now on Hugging Face under the MIT License.

What It Does

VibeVoice-ASR-Streaming is built as a unified streaming ASR system that continuously answers two questions as audio arrives: who is speaking and what they are saying. According to Microsoft, the model supports three core capabilities:

• Streaming speaker-attributed transcription — continuous transcription with speaker labels as speech streams in, rather than requiring a full audio clip upfront.
• Customized hotwords — users can supply domain-specific terms, names, or technical vocabulary to improve recognition accuracy on specialized content.
• Multilingual support — the model covers 10 languages: Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.

Specs and Availability

The Hugging Face model card lists the model at 9B parameters with BF16 tensor precision and a chat template, despite the "7B" designation in its name — a discrepancy Microsoft has not addressed in the model card. The repository has logged 889 downloads over the past month. As of publication, no inference provider has deployed the model for hosted API access, meaning developers must run it themselves via the accompanying GitHub repository.

Microsoft has published an accompanying technical report (arXiv 2609.02812) alongside the release, though the model card does not list specific benchmark scores such as word error rate (WER) figures, leaving accuracy claims to be verified through the linked evaluation materials and technical report rather than confirmed numbers on the model page itself.

The model is part of Microsoft's broader VibeVoice collection, described as "Frontier Text-to-Speech Models," suggesting the ASR-Streaming variant is a complementary speech-understanding counterpart to Microsoft's text-to-speech work in the same family. A live demo and full usage instructions are available through Microsoft's VibeVoice GitHub repository.

Pricing is not applicable in the traditional sense — the model is open-weight and self-hostable, with no per-token API pricing disclosed since no inference provider currently offers it as a hosted endpoint.

What This Means

Streaming ASR with built-in speaker diarization has typically required stitching together separate transcription and diarization pipelines, adding latency and complexity for use cases like live meeting transcription, call center analytics, or multilingual customer support. By unifying both tasks in a single streaming model with hotword customization, Microsoft is targeting production speech applications where domain vocabulary (product names, technical jargon) and speaker tracking matter simultaneously.

The lack of published benchmark numbers on the model card and the absence of hosted inference access are notable gaps — teams evaluating this model will need to run their own tests to confirm accuracy claims and will need to self-host for now. The 7B/9B parameter naming inconsistency is a minor but real red flag worth clarifying before production adoption. Still, an MIT-licensed, multilingual streaming ASR model from a major lab lowers the barrier for teams building real-time transcription products without relying on closed API providers.