---
title: VibeVoice ASR BitNet
emoji: 🎙️
colorFrom: blue
colorTo: yellow
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
python_version: "3.10"
pinned: false
license: mit
short_description: Multilingual CPU-only ASR with a 1.58-bit BitNet decoder
startup_duration_timeout: 1h
models:
  - microsoft/VibeVoice-ASR-BitNet
---

# VibeVoice-ASR-BitNet

Multilingual automatic speech recognition running **entirely on CPU** with
[`microsoft/VibeVoice-ASR-BitNet`](https://huggingface.co/microsoft/VibeVoice-ASR-BitNet):
an I8_S quantized acoustic/semantic VAE speech tokenizer feeding a ternary
(I2_S, ~1.58 bit) BitNet Qwen2.5-1.5B decoder — 1.58 GB of weights in total.

The model has no PyTorch/CUDA implementation; the only inference path is the
authors' ggml-based [VibeASR.cpp](https://github.com/microsoft/VibeASR.cpp)
engine, so this Space clones and compiles `asr_infer` from source on startup and
drives it as a subprocess. It runs entirely on CPU — CPU efficiency is the whole
point of the paper.

Languages: English, Chinese, French, Italian, Korean, Portuguese, Vietnamese.

Example clips come from LibriSpeech (English) and FLEURS (French, Mandarin).

Paper: [VibeVoice-ASR-BitNet Technical Report](https://huggingface.co/papers/2607.21075)