whatmodel.sh / stt / VibeVoice-ASR-BitNet

VibeVoice-ASR-BitNet

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

VibeVoice-ASR-BitNet is an open-weights speech-to-text model from Microsoft, released July 2026. Its published size is 2.81B parameters. Licence: mit.

Catalogue identity notes

Exact release microsoft/VibeVoice-ASR-BitNet @ 66e7802 (MIT, created 2026-07-24): 'a compressed variant of VibeVoice-ASR' for CPU inference with a Qwen2.5-1.5B LM (VibeASR.cpp README). Not the 7B VibeVoice-ASR (vibevoice-asr-hf) and not the streaming models.

Model information sources

Release date ↗ · Weights licence ↗ · Parameter count ↗

HF safetensors metadata: 2,814,116,321 F32 parameters in the original (pre-quantization) safetensors; VibeASR.cpp README: the Qwen2.5-7B LM of VibeVoice-ASR is replaced with Qwen2.5-1.5B. The released GGUF pair is I8_S (VAE tokenizer) + I2_S/Q6_K (LM).

50.2±4.0#98 of 162 speech-to-text modelsfrom 5 benchmark results across 1 source

Score within the Open ASR English short-form WER (March 2026) comparison set only.

$ how it compares

The speech-to-text models ranked either side of VibeVoice-ASR-BitNet, out of 176 with a score, then more from Microsoft.

#modelscorewhereRAMreleased
107Omnilingual ASR LLM 300M v2 Meta51±3local~4.4GB2025-12
108Moonshine Streaming Medium Useful Sensors (Moonshine AI)50±3local~1.8GB2026-01
109VibeVoice-ASR-BitNet Microsoft50±4local~1.7GB2026-07
110Chirp Google50±4comparison only2023-05
111Parakeet TDT-CTC 110M NVIDIA50±3local~1.1GB2024-09
more from Microsoft
57Phi-4-multimodal-instruct Microsoft58±2local~14GB2025-02
102VibeVoice-ASR-HF Microsoft52±3local~21GB2026-03
125VibeVoice-ASR-Streaming-7B Microsoft47±5local~17GB2026-09

see all speech-to-text models · find the ones that fit your machine

$ what it's good at

other tests5 not in a task · 94% of its score
testresultvs the others in its rundifficultycountswhy
LibriSpeech test-other WER lab6.3% ✎███·····██·█ beats 1/5avg 5.7%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 41%typical
VoxPopuli English test WER lab5.2% ✎▄··▄··▄···█▄ beats 4/5avg 7.4%▇▇▇▇▇▇▇▇ 19%typical
LibriSpeech test-clean WER lab2.4% ✎█·█·█·█···██ beats 1/5avg 2.2%▇▇▇▇▇ 12%blunt
FLEURS Mandarin test error rate lab8.3% ✎▄·▂▄▂▄██▂▂·▆ beats 12/20avg 7.3%▇▇▇▇▇ 11%typical
FLEURS English test WER lab5.2% ✎▄·····▄▄▄··█ beats 1/5avg 5.5%▇▇▇▇ 11%blunt

$ where to run it

Other ways to run it

  1. Run it as published · needs ~1.7 GB · from github.com/microsoft (official)
    git clone --recursive https://github.com/microsoft/VibeASR.cpp.git && cd VibeASR.cpp && cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j && pip install huggingface_hub && huggingface-cli download microsoft/VibeVoice-ASR-BitNet vibeasr-vae-encoder-i8_s.gguf vibeasr-lm-i2_s-embed-q6_k.gguf --local-dir models/vibeasr && ./build/bin/asr_infer --vae-model models/vibeasr/vibeasr-vae-encoder-i8_s.gguf --lm-model models/vibeasr/vibeasr-lm-i2_s-embed-q6_k.gguf --audio input.wav -t 4

Official weights and documentation: huggingface.co/microsoft ↗

$ how the score was worked out · microsoft/VibeVoice-ASR-BitNe… 94%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
microsoft/VibeVoice-ASR-BitNet model card (Accuracy table) · lab-reported5▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 94%MIT
prior · pull toward the middle▇▇ 6%
weighted average= 50.2

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json