Breaking

Microsoft Open-Sources VibeVoice-ASR-7B for Single-Pass Hour-Long Audio

June 9, 2026 at 00:01 EDT

Media Generation
Open Source
Foundation Models

In January 2026, Microsoft open-sourced VibeVoice-ASR-7B, a speech recognition model that processes up to 60 minutes of audio in a single pass and outputs structured transcription with speakers, timestamps and content, releasing it on GitHub. Model weights are distributed on Hugging Face under an MIT license, and local execution is possible.

January 21, 2026 · Microsoft

VibeVoice-ASR transcribes a full hour of audio in one pass

A free, locally-runnable speech model that handles up to 60 minutes of audio in a single inference — adding speaker labels and timestamps, with no 30-second chunking. Open weights, 50+ languages, structured JSON output.

60 min

audio in a single pass — no chunking

7.77%

average WER across 8 datasets

50+

languages, with code-switching

51.8×

RTFx — real-time speed factor

Single-pass context vs. chunked transcription

Each block is 30 seconds. Older systems slice audio into ~30s chunks, losing context and speaker tracking. VibeVoice-ASR reads it all at once.

~30s chunks

context resets every block

one continuous pass

60 min, single inference

7.5 Hz tokenizer · 64K context

Benchmark scorecard (lower is better, except RTFx)

LibriSpeech clean WER

2.20%

TED-LIUM WER

2.57%

DER (MLC-Challenge)

4.28%

Average WER (8 sets)

7.77%

cpWER (MLC-Challenge)

11.48%

What developers value

Free, runs locally — alternative to paid SaaS
Clean transcript with speaker labels + timestamps
Strong context retention on meetings, podcasts, lectures
Claimed open-source record on medical audio (WER 8.34%)

Limitations noted

~9B model needs ~18GB VRAM — won't run on a T4
CPU inference is heavier still
Lower diarization accuracy on utterances under 1s
Summary quality / ease of use not yet at SaaS level

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.