Microsoft open-sources VibeVoice-ASR for 60-minute single-pass audio

Microsoft has open-sourced VibeVoice-ASR, a 9B parameter speech understanding model designed specifically to fix context fragmentation and speaker confusion in long-form audio.

Most current speech recognition setups handle long recordings by chopping audio into 30-second chunks, running ASR, and stitching texts together. That usually wrecks context and makes speaker identification messy. VibeVoice-ASR drops the chunking pipeline entirely and processes up to 60 minutes of continuous audio in a single pass.

What it does

Instead of chaining multiple standalone tools, VibeVoice-ASR unifies automatic speech recognition, speaker diarization, and precise timestamping into a single end-to-end generation task. You feed it raw audio, and it outputs a clean transcript with timestamps and speaker tags already attached.

Speed and accuracy look very solid. Reports note it can transcribe two minutes of audio in under six seconds, delivering error rates well below larger baselines like Whisper.

According to its research paper, it supports over 50 languages without requiring manual language toggles, while other reports put coverage at over 100 languages. It natively handles code-switching between languages within the same sentence.

It also introduces a prompt-based context injection feature. Users can supply domain glossaries or context hints beforehand, which helps the model nail tricky technical terminology and disambiguate polyphonic characters.

Access and requirements

The research paper is up on arXiv (arXiv:2601.18184), and the codebase is open-sourced with full fine-tuning support.

Because this is a 9B parameter model, you will need serious local GPU hardware with substantial VRAM to run single-pass 60-minute inference or perform custom fine-tuning.

Why it matters

If you regularly deal with hour-long podcasts, all-hands meetings, or user interviews, VibeVoice-ASR is a massive upgrade over traditional multi-step pipelines. Slicing audio into chunks was always a fragile workaround, and consolidating ASR and diarization into one model cuts out endless pipeline friction.

In our view, single-pass processing is quickly becoming the benchmark for frontier speech models. While 9B parameters make it heavier to host locally than tiny edge models, it provides teams building private transcription systems a truly formidable open-source alternative to Whisper.