whatmodel.sh / stt / Cohere Transcribe (03-2026)
Cohere Transcribe (03-2026)
↗ share
+ contribute
Cohere Transcribe (03-2026) is an open-weights speech-to-text model from Cohere, released March 2026. Its published size is 2.07B parameters. Licence: apache-2.0.
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 (03-2026), out of 176 with a score, then more from Cohere.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 23 |  Inworld STT-1 | 62±6 | api $0.0025/audio min | ||
| 24 |  Qwen-Audio-3.0-ASR | 62±4 | comparison only | ||
| 25 | Cohere Transcribe (03-2026) | 62±2 | api $0.0042/audio min +local | ||
| 26 |  Granite Speech 4.1 2B NAR | 62±4 | local | ||
| 27 |  Nova-3 | 61±2 | api $0.0052/audio min | ||
| more from Cohere | |||||
| 16 | Cohere Transcribe Arabic (07-2026) | 63±3 | local | ||
see all speech-to-text models · find the ones that fit your machine
$ what it's good at
long-form transcription▁▁▄▁·▁▂█▄▃▁▁ #3 of 27 * · 11% of its score · 1 test
Trails Speechmatics Enhanced here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Open ASR English long-form WER | 7.9% ✎ | ▁▁▅▁·▃▅█▆▃▂▂ beats 26/30 | 11% |
multilingual transcription▂··▃▄█▆▅▄▂·▂ #6 of 36 * · 37% of its score · 1 test
Beats Cohere Transcribe Arabic (07-2026) here, though they rank higher overall. Trails SeamlessM4T v2 Large, Voxtral Small 24B (2507), Chirp 3 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Open ASR multilingual WER | 3.8% ✎ | ▂··▂▃▂█▃█▃▁▁ beats 26/29 | 37% |
other tests8 not in a task · 51% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LibriSpeech test-clean WER | 1.3% ✎ | ▁···▁▁▂·▄▄██ beats 56/56 | 14% |
| Open ASR English short-form WER (March 2026) | 5.4% ✎ | ▁▁▁▃··▁▁▂██▆ beats 86/86 | 11% |
| Dutch ASR Leaderboard · average WER (FLEURS, VoxPopuli, MLS nl) | 8.4% ✎ | ▄··▄█···█··▄ beats 4/6 | 8.0% |
| FLEURS English test WER | 5.1% ✎ | ▁··▁▁▁▁▁▁▃█▁ beats 24/56 | 5.6% |
| FLEURS Mandarin test error rate | 11.2% ✎ | ▂▂··▂▄▂▂·▆██ beats 7/17 | 5.4% |
| Open Universal Arabic ASR Leaderboard · 6-test-set average WER | 30.7% ✎ | ▃·▂▂▅█▅▃▂▅▂▂ beats 18/22 | 3.2% |
| Quranic ASR Leaderboard · held-out recitation (600 clips) overall WER | 15.2% ✎ | ▃·▅·▅▃▃·▃▃▃█ beats 7/12 | 2.0% |
| Speech Accessibility Project dysarthric speech · zero-shot WER (rated subset, N=11,218) | 14.5% ✎ | ▂·▂··▂····▆█ beats 6/9 | 1.3% |
$ where to run it
Through an API
api · fal.ai · $0.0042/audio min · fal.ai ↗ · price sourceUSD per input audio minute: fal prices Cohere Transcribe (fal-ai/cohere-transcribe) at $0.00006944444 per second (x60 = $0.0041667/min; endpoint billing record: billing_unit seconds). Cohere's docs give Cohere Transcribe's model name as cohere-transcribe-03-2026 (its only general Transcribe model; Cohere Transcribe Arabic is separate). Cohere's own API bills Transcribe per token, which is not converted. Product: fal speech-to-text endpoint. API model: fal-ai/cohere-transcribe. Settings: standard. Price basis: published list rate, per-second to per-minute scaling only. Currency: USD. fal.ai model page, observed 2026-09-25.
Other ways to run it
- Run it as published · needs ~5.47 GB, NVIDIA only · from huggingface.co/CohereLabs
vllm serve CohereLabs/cohere-transcribe-03-2026 --trust-remote-code
Weights and documentation: huggingface.co/CohereLabs ↗
$ how the score was worked out · Open ASR Leaderboard 59%, transcribe.cpp model catalog… 25%, Dutch ASR Leaderboard (tvosch… 8% and 3 more
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
| source | results | share of the score |
|---|---|---|
 Open ASR Leaderboard | 3 | 59% |
 transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports | 3 | 25% |
| Dutch ASR Leaderboard (tvosch) · committed result files | 1 | 8% |
| Open Universal Arabic ASR Leaderboard (Elm Research Center) | 1 | 3% |
| Quranic ASR Leaderboard (Quran-Lab, held-out recitation set) | 1 | 2% |
 Dysarthric ASR with clinical context on the Speech Accessibility Project (arXiv:2605.02782v2) | 1 | 1% |
| prior · pull toward the middle | 1% | |
| weighted average | = 61.7 |
$ improve this page
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