whatmodel.sh / stt / Breeze ASR 25

Breeze ASR 25

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

Breeze ASR 25 is an open-weights speech-to-text model from MediaTek Research, released June 2025. Its published size is 1.543B parameters. Licence: apache-2.0.

Catalogue identity notes

MediaTek-Research/Breeze-ASR-25, fine-tuned from Whisper-large-v2 for Taiwanese Mandarin and Mandarin-English code-switching (card front matter base_model openai/whisper-large-v2). Not whisper-large-v2 (catalogued). transcribe.cpp port pins upstream cffe7cc (2025-07-08, current main); repo created 2025-06-06.

Model information sources

Release date ↗ · Parameter count ↗

Hugging Face safetensors metadata total 1,543,304,960 (current main cffe7cc); transcribe.cpp catalog counts 1,543,321,440 for the converted port.

54.4±3.5#78 of 162 speech-to-text modelsfrom 3 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 Breeze ASR 25, out of 176 with a score.

#modelscorewhereRAMreleased
87Qwen3-ASR-0.6B Alibaba (Qwen)55±1local~2.8GB2026-01
88Ink-Whisper Cartesia55±3api price unknown2025-06
89Breeze ASR 25 MediaTek Research54±4local~1.7GB2025-06
90Parakeet CTC 0.6B NVIDIA54±3local~3.4GB2024-01
91Canary-180M-Flash NVIDIA54±3local~1.4GB2025-03

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

$ what it's good at

other tests3 not in a task · 95% of its score
testresultvs the others in its rundifficultycountswhy
LibriSpeech test-clean WER2.3% ✎▁···▁▁▂·▄▄██ beats 19/56avg 2.3%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 54%blunt
FLEURS English test WER4.1% ✎▁··▁▁▁▁▁▁▃█▁ beats 48/56avg 5.8%▇▇▇▇▇▇▇▇▇ 21%blunt
FLEURS Mandarin test error rate8.1% ✎▂▂··▂▄▂▂·▆██ beats 14/17avg 12%▇▇▇▇▇▇▇▇ 20%typical

$ where to run it

Other ways to run it

  1. Run it as published · needs ~1.67 GB · from github.com/handy-computer
    git clone https://github.com/handy-computer/transcribe.cpp.git && cd transcribe.cpp && cmake -B build && cmake --build build && hf download handy-computer/Breeze-ASR-25-gguf Breeze-ASR-25-Q8_0.gguf --local-dir models/breeze-asr-25 && build/bin/transcribe-cli -m models/breeze-asr-25/Breeze-ASR-25-Q8_0.gguf audio.wav

Official weights and documentation: huggingface.co/MediaTek-Research ↗

$ how the score was worked out · transcribe.cpp model catalog… 95%

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
transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports3▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 95%MIT
prior · pull toward the middle▇ 5%
weighted average= 54.4

$ 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