whatmodel.sh / stt / Granite Speech 4.1 2B NAR

Granite Speech 4.1 2B NAR

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

Granite Speech 4.1 2B NAR is an open-weights speech-to-text model from IBM, released March 2026. Its published size is 2.255B parameters. Licence: apache-2.0.

Catalogue identity notes

ibm-granite/granite-speech-4.1-2b-nar: non-autoregressive sibling of granite-speech-4-1-2b (card: 'formulates ASR as conditional transcript editing', NLE architecture, arXiv:2603.08397). Not the autoregressive granite-speech-4-1-2b (catalogued) or the -plus model. transcribe.cpp port pins upstream 99a4df9 (2026-05-21); only later commit a1e3416 edits README.md. Repo created 2026-03-10.

Model information sources

Release date ↗ · Parameter count ↗

Hugging Face safetensors metadata total 2,254,635,356 (current main a1e3416); transcribe.cpp catalog counts 2,254,656,316 for the converted port.

61.5±3.9#22 of 162 speech-to-text modelsfrom 2 benchmark results across 1 source* thinly tested, treat the rank as provisional

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 Granite Speech 4.1 2B NAR, out of 176 with a score, then more from IBM.

#modelscorewhereRAMreleased
24Qwen-Audio-3.0-ASR Alibaba62±4comparison only2026-09
25Cohere Transcribe (03-2026) Cohere62±2api $0.0042/audio min +local~5.5GB2026-03
26Granite Speech 4.1 2B NAR IBM62±4local~2.5GB2026-03
27Nova-3 Deepgram61±2api $0.0052/audio min2025-02
28Canary-Qwen-2.5B NVIDIA61±2local~6.6GB2025-07
more from IBM
15Granite Speech 5.0 470M TurboCTC NC IBM64±4local~0.5GB2026-08
18Granite Speech 4.1 2B IBM62±3local~6GB2026-04
19Granite 4.0 1B Speech IBM62±3local~6GB2026-03

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

$ what it's good at

other tests2 not in a task · 94% of its score
testresultvs the others in its rundifficultycountswhy
LibriSpeech test-clean WER1.3% ✎▁···▁▁▂·▄▄██ beats 55/56avg 2.3%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 67%blunt
FLEURS English test WER5.3% ✎▁··▁▁▁▁▁▁▃█▁ beats 22/56avg 5.8%▇▇▇▇▇▇▇▇▇▇▇ 27%blunt

$ where to run it

Other ways to run it

  1. Run it as published · needs ~2.5 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/granite-speech-4.1-2b-nar-gguf granite-speech-4.1-2b-nar-Q8_0.gguf --local-dir models/granite-speech-4.1-2b-nar && build/bin/transcribe-cli -m models/granite-speech-4.1-2b-nar/granite-speech-4.1-2b-nar-Q8_0.gguf audio.wav

Weights and documentation: huggingface.co/ibm-granite ↗

$ how the score was worked out · transcribe.cpp model catalog… 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
transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports2▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 94%MIT
prior · pull toward the middle▇▇ 6%
weighted average= 61.5

$ 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