TEST RECORD · ← all field reports

FIELD REPORTS/AUDIO/ISSUE #85

Transcribe speech in seven languages in real time

Offline speech recognition used to be a trade: either accept toy accuracy or find a machine with a serious graphics card. Microsoft compressed its VibeVoice transcriber until it runs in real time on an ordinary processor, and the accuracy mostly survived the squeeze.

MODELVibeVoice-ASR-BitNet
PUBLISHEDAugust 7, 2026
READ TIME2 min
TESTED BYNeural Expedition
CATEGORYAUDIO

Field notes

01What it does

You point it at a recording or a microphone and it types out what was said, in English, Chinese, French, Italian, Korean, Portuguese, or Vietnamese. The whole model fits in about 1.6 gigabytes and keeps up with live audio using three or four threads of a normal laptop processor, no graphics card involved. Nothing leaves your machine, which is the point: the audio of a client call or a doctor's dictation never touches a server.

The trick is aggressive compression. Microsoft stored most of the model's weights at close to two bits each, a technique called BitNet, then wrote custom code so ordinary processors can chew through those tiny weights quickly. The result runs 1.6 to 2.3 times faster than Whisper.cpp, the current default for local transcription, and on messy meeting recordings it also makes fewer mistakes than Whisper.

02How to try it

Start with the official demo Space: record or upload a clip of a real conversation, ideally one with cross-talk or background noise, and compare the transcript against what you would expect from your current tool. Meeting-style audio is where this model beats Whisper most clearly, so that is the test worth running. For daily use, the VibeASR.cpp project on GitHub builds on Mac, Windows, and Linux, and the two model files download straight from the Hugging Face page.

03Caveat

Compression costs a little accuracy: expect a few more errors per hundred words than the full-size VibeVoice model, with the biggest drop on Chinese meeting recordings. Crowded rooms with several people talking over each other remain the hardest case for it, as for every transcriber. Clean single-speaker audio in English is where it is essentially as good as the original.

04What you can do with it

  • Take meeting notes on a work laptop with the audio never leaving the machine.
  • Transcribe interviews in any of the seven languages without a cloud account.
  • Batch-transcribe a podcast or lecture archive overnight on a desktop.
  • Add live captions to a kiosk or a small single-board computer.
  • Dictate notes offline on the train and keep them private.

Try the demo

View model page

Read this issue on neuralexpedition.com →