Microsoft's VibeVoice Streaming ASR Models Go Open Weights
Microsoft released open weights for VibeVoice-ASR-Streaming in 7B and 1.5B sizes: speech-to-text models that turn up to 60 minutes of audio into a single structured transcription with speakers, timestamps, and content in one pass. The weights are on Hugging Face, community GGUF builds exist, and the 7B model needs roughly 24 GB of VRAM.
On this page
What the models do differently
Microsoft's VibeVoice-ASR reformulates transcription as a generation problem. Instead of outputting plain text, the model produces what the technical report calls a Rich Transcription: a structured stream that interleaves speaker identities, timestamps, and content, unifying speech recognition, speaker diarization, and timing in a single pass [1]. The report claims the lowest average word and character error rates across five evaluation sets, and support for more than 50 languages without an explicit language setting [1].
The new weights published on Hugging Face add streaming: the VibeVoice-ASR-Streaming-7B and Streaming-1.5B variants were updated hours ago, and both are listed in Microsoft's official VibeVoice collection [3]. The underlying system handles 60 minutes of continuous audio in a single pass through a 64K-token context window, with streaming segmentation for anything longer [1].
What shipped, and for which hardware
The open-weights release lives in the microsoft/VibeVoice repository, described as open-source frontier voice AI, alongside TTS and realtime models [2]. The ASR family now spans the full-size VibeVoice-ASR, the new Streaming-7B, a smaller Streaming-1.5B for lighter machines, and an earlier BitNet variant [3].
Hardware is the honest caveat: the 7B model wants roughly 24 GB of VRAM, which puts it in workstation or datacenter GPU territory rather than entry-level laptops [1]. The 1.5B streaming variant is the realistic option for more modest machines, and community quantizations, including GGUF builds, already exist for people who want to trade some accuracy for a smaller footprint [3].
Why local transcription matters
Running speech-to-text locally removes the two biggest reasons people avoid it for sensitive audio. Nothing about a private meeting, an interview, or a medical conversation needs to leave the machine, and there are no per-minute API charges accumulating over a long recording [1]. The single-pass design also removes the usual stitching layer, where separate diarization, timestamping, and transcription tools disagree with each other [1].
The tooling ecosystem is already forming. Community GGUF quantizations are circulating for local runtimes [3], the models are integrated into Hugging Face Transformers, and the repository bundles ASR with the existing VibeVoice text-to-speech and realtime models [2]. For anyone building a local note-taking or meeting pipeline, that is most of the pieces arriving at once [2].
What to weigh before running it
Accuracy claims come from the authors' own evaluation, so independent verification is still worth waiting for, and the community has already flagged minor issues with short interjections [1]. The VRAM requirement makes the 7B model a poor fit for older machines even after quantization [1]. But as a direction, fresh open weights for streaming, speaker-aware transcription are exactly the kind of release that keeps sensitive audio work feasible on local hardware [3].