whatmodel.sh / stt / Canary-1B-v2
Canary-1B-v2
↗ share
+ contribute
Canary-1B-v2 is an open-weights speech-to-text model from NVIDIA, released August 2025. Its published size is 0.98B 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-1B-v2, out of 176 with a score, then more from NVIDIA.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 54 | Parakeet RNNT 1.1B | 58±2 | local | ||
| 55 |  Chirp 2 | 58±3 | api price unknown | ||
| 56 | Canary-1B-v2 | 58±1 | local | ||
| 57 |  Phi-4-multimodal-instruct | 58±2 | local | ||
| 58 |  Omnilingual ASR LLM 7B | 58±2 | local | ||
| more from NVIDIA | |||||
| 28 | Canary-Qwen-2.5B | 61±2 | local | ||
| 35 | Parakeet TDT 0.6B v2 | 60±1 | local | ||
| 37 | Parakeet Unified EN 0.6B | 60±4 | local | ||
see all speech-to-text models · find the ones that fit your machine
$ what it's good at
multilingual transcription▂··▃▄█▆▅▄▂·▂ #25 of 36 * · 22% of its score · 2 tests
Beats Voxtral Mini Transcribe V2 here, though they rank higher overall. Trails Omnilingual ASR CTC 1B v2, Omnilingual ASR CTC 7B v2, SeamlessM4T v2 Large here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Open ASR multilingual WER | 4.6% ✎ | ▂··▂▃▂█▃█▂▁▁ beats 22/29 | 21% |
| Iberian ASR benchmark · 7-language macro WER (eu, ca, gl, pt, es + de, tr) | 62.1% ✎ | █▃▃··▅▃▃··▃▃ beats 0/10 | 1.1% |
other tests17 not in a task · 77% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LibriSpeech test-clean WER +1 | 1.9% ✎ | ▁···▁▁▂·▄▄██ beats 29/56 | 9.2% |
| Mega-ASR adverse-condition suite · mean WER (CHiME-4, VOiCES, NOIZEUS) | 11.8% ✎ | ▄····█▄▄···▄ beats 2/5 | 6.9% |
| SPGISpeech test WER lab | 2.3% ✎ | ▄▄··▄██··▄██ beats 7/11 | 6.4% |
| FLEURS 25 European languages, mean WER (NVIDIA set) lab +1 | 8.4% ✎ | ██·········█ beats 2/2 | 6.3% |
| Open ASR English short-form WER (March 2026) | 7.1% ✎ | ▁▁▁▃··▁▁▂██▆ beats 55/86 | 6.3% |
| CoVoST 2 source speech, 13 languages, mean ASR WER (NVIDIA set) lab +1 | 8.9% ✎ | █·····█····█ beats 2/2 | 6.2% |
| Multilingual LibriSpeech 6 languages, mean WER (NVIDIA set) lab +1 | 7.3% ✎ | █·▄········▄ beats 2/3 | 5.8% |
| Earnings-22 WER (short-form, Open ASR / ESB test split) lab | 11.8% ✎ | █······███·█ beats 1/4 | 5.3% |
9 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| AMI WER (Open ASR / ESB test split) lab | 16.0% ✎ | ▄▄·······█·▄ beats 1/4 | 5.0% |
| GigaSpeech test WER lab | 10.8% ✎ | ▁····▁··▁▄█▂ beats 3/17 | 4.6% |
| LibriSpeech test-other WER lab | 3.6% ✎ | ▁▁▁▂▅▃▅▄█▃▂▁ beats 48/54 | 3.8% |
| FLEURS English test WER | 4.5% ✎ | ▁··▁▁▁▁▁▁▃█▁ beats 36/56 | 3.2% |
| Open ASR Leaderboard – Chilean Spanish · 3-dataset average WER | 9.8% ✎ | ▂▂·····▃▂·█▂ beats 2/8 | 3.0% |
| VoxPopuli English test WER lab | 6.3% ✎ | █▄▄··████▄██ beats 10/16 | 2.5% |
| Danish ASR Leaderboard · mean WER over 5 Danish test sets (RyeAI) | 24.7% ✎ | ▂▂···▃▂▅█▃▂▂ beats 11/16 | 2.4% |
| TED-LIUM test WER (short-form) lab | 4.3% ✎ | ▁·▁▇█▃·▁▁▃█▄ beats 18/27 | 0.2% |
| Earnings-22 long-form WER lab | 13.8% ✎ | █······█·███ beats 3/4 | — |
$ where to run it
Other ways to run it
version
- Run it as published · needs ~5.2 GB · from github.com/istupakov
uvx --from 'onnx-asr[cpu,hub]' onnx-asr nemo-canary-1b-v2 audio.wav
Official weights and documentation: huggingface.co/nvidia ↗
$ how the score was worked out · NVIDIA ASR model cards (Parak… 36%, Open ASR Leaderboard 27%, transcribe.cpp model catalog… 11% and 5 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 | 11 | 36% |
| Open ASR Leaderboard | 2 | 27% |
 transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports | 2 | 11% |
 Canary-1B-v2 & Parakeet-TDT-0.6B-v3: Efficient and High-Performance Models for Multilingual ASR and AST (NVIDIA, arXiv:2509.14128v2) · lab-reported | 3 | 11% |
| Mega-ASR (arXiv:2605.19833v1) · adverse-condition ASR baselines (CHiME-4, VOiCES, NOIZEUS) | 1 | 7% |
| Open ASR Leaderboard – Chilean Spanish (idsudd / Alonso Astroza) | 1 | 3% |
| Danish Speech-to-Text Leaderboard (RyeAI) · results dataset | 1 | 2% |
| Benchmarking Automatic Speech Recognition Tools for Iberian Languages (arXiv:2609.36920v1), Table 3 | 1 | 1% |
| prior · pull toward the middle | 1% | |
| weighted average | = 57.9 |
not counted (1)
| NVIDIA ASR model cards (Parakeet / Canary, Hugging Face) | Earnings-22 long-form WER [correct] | 13.8% | fewer than three models have this test yet, or this source ran it for this model alone |
$ published speed measurements · 4
Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.
| AMD Ryzen 7 9800X3D | onnx-default (precision not stated in table) | 8.3 x-realtime | ONNX Runtime CPUExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
| AMD Ryzen 7 9800X3D | onnx-int8 | 14.9 x-realtime | ONNX Runtime CPUExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
| NVIDIA GeForce RTX 5070 Ti | onnx-default (precision not stated in table) | 35.7 x-realtime | ONNX Runtime CUDAExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
| NVIDIA T4 | onnx-default (precision not stated in table) | 21.4 x-realtime | ONNX Runtime CUDAExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
$ 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