whatmodel.sh / stt / Cohere Transcribe Arabic (07-2026)

Cohere Transcribe Arabic (07-2026)

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

Cohere Transcribe Arabic (07-2026) is an open-weights speech-to-text model from Cohere, released July 2026. Its published size is 2.07B parameters. Licence: apache-2.0.

Catalogue identity notes

Exact checkpoint CohereLabs/cohere-transcribe-arabic-07-2026 (gated HF repo, Apache-2.0), Cohere's Arabic-focused Transcribe model; distinct from the general cohere-transcribe-03-2026. Params: HF safetensors total 2,065,804,048. released: the model name's 07-2026 version date (HF repo created 2026-06-18). No local command added (gated repo; not verified here).

63.2±3.2#14 of 162 speech-to-text modelsfrom 4 benchmark results across 3 sources

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 Cohere Transcribe Arabic (07-2026), out of 176 with a score, then more from Cohere.

#modelscorewhereRAMreleased
14Zoom Scribe v1 Zoom64±5api price unknown2026-03
15Granite Speech 5.0 470M TurboCTC NC IBM64±4local~0.5GB2026-08
16Cohere Transcribe Arabic (07-2026) Cohere63±3local~4.1GB2026-07
17Universal-3 Pro AssemblyAI63±2comparison only2026-02
18Granite Speech 4.1 2B IBM62±3local~6GB2026-04
more from Cohere
25Cohere Transcribe (03-2026) Cohere62±2api $0.0042/audio min +local~5.5GB2026-03

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

$ what it's good at

multilingual transcription▂··▃▄█▆▅▄▂·▂ #16 of 36 * · 21% of its score · 1 test

Trails Omnilingual ASR CTC 1B v2, SeamlessM4T v2 Large, Omnilingual ASR LLM 1B v2 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
QCRI Arabic-English ASR leaderboard, 22-set average WER20.9% ✎█···█······█ beats 2/2avg 53%▇▇▇▇▇▇▇▇ 21%small run
other tests3 not in a task · 75% of its score
testresultvs the others in its rundifficultycountswhy
LibriSpeech test-clean WER1.3% ✎▁···▁▁▂·▄▄██ beats 52/56avg 2.3%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 46%blunt
FLEURS English test WER4.9% ✎▁··▁▁▁▁▁▁▃█▁ beats 26/56avg 5.8%▇▇▇▇▇▇▇ 18%blunt
Open Universal Arabic ASR Leaderboard · 6-test-set average WER25.9% ✎▃·▂▂▅█▅▃▂▅▂▃ beats 22/22avg 40%▇▇▇▇ 10%blunt

$ where to run it

Other ways to run it

  1. Run it as published · needs ~4.13 GB · from huggingface.co/CohereLabs
    uv run --with 'transformers>=5.4.0' --with torch --with huggingface_hub --with soundfile --with librosa --with sentencepiece --with protobuf --with accelerate python -c "from transformers import AutoProcessor, CohereAsrForConditionalGeneration; from transformers.audio_utils import load_audio; processor = AutoProcessor.from_pretrained('CohereLabs/cohere-transcribe-arabic-07-2026'); model = CohereAsrForConditionalGeneration.from_pretrained('CohereLabs/cohere-transcribe-arabic-07-2026', device_map='auto'); audio = load_audio('audio.wav', sampling_rate=16000); inputs = processor(audio, sampling_rate=16000, return_tensors='pt', language='ar'); inputs.to(model.device, dtype=model.dtype); outputs = model.generate(**inputs, max_new_tokens=256); print(processor.decode(outputs, skip_special_tokens=True))"

Weights and documentation: huggingface.co/CohereLabs ↗

$ how the score was worked out · transcribe.cpp model catalog… 65%, QCRI Arabic + English ASR lea… 21%, Open Universal Arabic ASR Lea… 10%

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▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 65%MIT
QCRI Arabic + English ASR leaderboard (Hugging Face Space QCRI/AURA-Arabic-English-ASR-leaderboard)1▇▇▇▇▇▇ 21%CC BY 4.0
Open Universal Arabic ASR Leaderboard (Elm Research Center)1▇▇▇ 10%Apache-2.0
prior · pull toward the middle▇ 4%
weighted average= 63.2

$ 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