whatmodel.sh / stt / Canary-Qwen-2.5B
Canary-Qwen-2.5B
↗ share
+ contribute
Canary-Qwen-2.5B is an open-weights speech-to-text model from NVIDIA, released July 2025. Its published size is 2.56B parameters. Licence: cc-by-4.0.
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 Canary-Qwen-2.5B, out of 176 with a score, then more from NVIDIA.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 26 |  Granite Speech 4.1 2B NAR | 62±4 | local | ||
| 27 |  Nova-3 | 61±2 | api $0.0052/audio min | ||
| 28 | Canary-Qwen-2.5B | 61±2 | local | ||
| 29 |  Pulse (STT) | 61±3 | api $0.0040/audio min | ||
| 30 |  Scribe v1 | 61±2 | comparison only | ||
| more from NVIDIA | |||||
| 35 | Parakeet TDT 0.6B v2 | 60±1 | local | ||
| 37 | Parakeet Unified EN 0.6B | 60±4 | local | ||
| 40 | Canary-1B-Flash | 60±2 | local | ||
see all speech-to-text models · find the ones that fit your machine
$ what it's good at
long-form transcription▁▁▄▁·▁▂█▄▁▁▁ #4 of 27 * · 7.4% 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 | 8.6% ✎ | ▁▁▅▁·▃▅█▆▁▂▂ beats 24/30 | 7.4% |
other tests14 not in a task · 92% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| AfriSpeech-MultiBench · 7-region African-accent mean WER | 22.4% ✎ | ▂····▂···▂▆█ beats 9/11 | 14% |
| AfriSpeech-MultiBench · Parliament (legal domain) WER | 18.2% ✎ | ▂▂····▂·▅·▃█ beats 11/12 | 12% |
| LibriSpeech test-clean WER +1 | 1.6% ✎ | ▁···▁▁▂·▄▄██ beats 40/56 | 11% |
| SPGISpeech test WER lab | 1.9% ✎ | ▄▄··▄██··▄██ beats 10/11 | 7.5% |
| Open ASR English short-form WER (March 2026) | 5.6% ✎ | ▁▁▁▃··▁▁▂██▆ beats 83/86 | 7.4% |
| Earnings-22 WER (short-form, Open ASR / ESB test split) lab | 10.4% ✎ | █······███·█ beats 4/4 | 6.2% |
| AMI WER (Open ASR / ESB test split) lab | 10.2% ✎ | ▄▄·······█·▄ beats 4/4 | 5.9% |
| RW-Voice-EQ ASR · 4 real-world English sets, mean WER | 9.2% ✎ | █▃▃▅▃···▃··▃ beats 3/9 | 5.8% |
6 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| GigaSpeech test WER lab | 9.4% ✎ | ▁····▁··▁▄█▂ beats 15/17 | 5.5% |
| AfriCodeSwitch English-Yoruba code-switched speech · WER (cap2000) | 70.0% ✎ | █··█·█···█·█ beats 2/4 | 5.5% |
| LibriSpeech test-other WER lab | 3.1% ✎ | ▁▁▁▂▅▃▅▄█▂▃▁ beats 53/54 | 4.5% |
| FLEURS English test WER | 3.6% ✎ | ▁··▁▁▁▁▁▁▃█▃ beats 54/56 | 3.8% |
| VoxPopuli English test WER lab | 5.7% ✎ | █▄▄··████▄██ beats 12/16 | 3.0% |
| TED-LIUM test WER (short-form) lab | 2.7% ✎ | ▁·▁▇█▃·▁▁▃█▄ beats 27/27 | 0.3% |
$ where to run it
Other ways to run it
- Run it as published · needs ~6.64 GB, NVIDIA only · from huggingface.co/nvidia (official)
python examples/speechlm2/salm_generate.py pretrained_name=nvidia/canary-qwen-2.5b inputs=input_manifest.json output_manifest=generations.jsonl batch_size=128 user_prompt="Transcribe the following:"
Official weights and documentation: huggingface.co/nvidia ↗
$ how the score was worked out · NVIDIA ASR model cards (Parak… 34%, AfriSpeech-MultiBench (arXiv:… 26%, Open ASR Leaderboard 15% 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 |
|---|---|---|
 NVIDIA ASR model cards (Parakeet / Canary, Hugging Face) · lab-reported | 8 | 34% |
 AfriSpeech-MultiBench (arXiv:2511.14255v1) · African-accented English ASR benchmark suite | 2 | 26% |
| Open ASR Leaderboard | 2 | 15% |
 transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports | 2 | 13% |
| RW-Voice-EQ Bench (Hume AI) | 1 | 6% |
| Switch-aware evaluation of ASR on English-Yoruba code-switched speech (arXiv:2609.11786v1) | 1 | 5% |
| prior · pull toward the middle | 1% | |
| weighted average | = 61.2 |
$ 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