Official documentation
canary-1b-v2 model card (raw README)
NVIDIA, Hugging Face (NVIDIA model card), 2025
Read the originalRead on October 19, 2026
- What it is
- NVIDIA's model card for Canary 1B v2, read from its raw README. It gives average word error rate over the FLEURS, CoVoST and Multilingual LibriSpeech languages, the English Open ASR Leaderboard datasets, noise robustness, and long-form recordings.
- How much weight it carries
- Self-reported by NVIDIA, the model developer. The card uses its own text normalization and decoding, so its figures are not directly comparable with other vendors' full-transcript scores. Compare only within the same test set.
Facts taken from this source
- NVIDIA Canary 1B v2 had a word error rate of 16.01 percent on the AMI meetings test set in the Open ASR Leaderboard table.
“| **WER ↓** | **RTFx** | **Mean** | **AMI** | **GigaSpeech** | **LS Clean** | **LS Other** | **Earnings22** | **SPGISpech** | **Tedlium** | **Voxpopuli** | ... | `Canary-1b-v2` | 749 | 7.15 | 16.01 | 10.82 | 2.18 | 3.56 | 11.79 | 2.28 | 4.29 | 6.25 |” (Open ASR Leaderboard table, AMI meetings column)
- NVIDIA Canary 1B v2 had an average word error rate of 8.85 percent on the CoVoST 2 test across its CoVoST-13 languages.
“| **WER ↓** | Fleurs-25 Langs | CoVoST-13 Langs | MLS - 6 Langs | ... | **`Canary-1b-v2`** | 8.40% | 8.85% | 7.27% |” (ASR table, CoVoST-13 column)
- NVIDIA Canary 1B v2 had a word error rate of 13.78 percent on the Earnings-22 long-form long-form recordings.
“| **Dataset** | **WER ↓** | ... | Earnings-22 | 13.78% |” (Long-form inference table, Earnings-22 row)
- NVIDIA Canary 1B v2 had a word error rate of 11.79 percent on the Earnings-22 test set in the Open ASR Leaderboard table.
“| **WER ↓** | **RTFx** | **Mean** | **AMI** | **GigaSpeech** | **LS Clean** | **LS Other** | **Earnings22** | **SPGISpech** | **Tedlium** | **Voxpopuli** | ... | `Canary-1b-v2` | 749 | 7.15 | 16.01 | 10.82 | 2.18 | 3.56 | 11.79 | 2.28 | 4.29 | 6.25 |” (Open ASR Leaderboard table, Earnings-22 column)
- NVIDIA Canary 1B v2 had an average word error rate of 8.40 percent on the FLEURS test across its Fleurs-25 languages.
“| **WER ↓** | Fleurs-25 Langs | CoVoST-13 Langs | MLS - 6 Langs | ... | **`Canary-1b-v2`** | 8.40% | 8.85% | 7.27% |” (ASR table, Fleurs-25 column)
- NVIDIA Canary 1B v2 had a word error rate of 10.82 percent on the GigaSpeech test set in the Open ASR Leaderboard table.
“| **WER ↓** | **RTFx** | **Mean** | **AMI** | **GigaSpeech** | **LS Clean** | **LS Other** | **Earnings22** | **SPGISpech** | **Tedlium** | **Voxpopuli** | ... | `Canary-1b-v2` | 749 | 7.15 | 16.01 | 10.82 | 2.18 | 3.56 | 11.79 | 2.28 | 4.29 | 6.25 |” (Open ASR Leaderboard table, GigaSpeech column)
- NVIDIA Canary 1B v2 had a word error rate of 2.18 percent on the LibriSpeech test-clean test set in the Open ASR Leaderboard table.
“| **WER ↓** | **RTFx** | **Mean** | **AMI** | **GigaSpeech** | **LS Clean** | **LS Other** | **Earnings22** | **SPGISpech** | **Tedlium** | **Voxpopuli** | ... | `Canary-1b-v2` | 749 | 7.15 | 16.01 | 10.82 | 2.18 | 3.56 | 11.79 | 2.28 | 4.29 | 6.25 |” (Open ASR Leaderboard table, LibriSpeech test-clean column)
- NVIDIA Canary 1B v2 had a word error rate of 3.56 percent on the LibriSpeech test-other test set in the Open ASR Leaderboard table.
“| **WER ↓** | **RTFx** | **Mean** | **AMI** | **GigaSpeech** | **LS Clean** | **LS Other** | **Earnings22** | **SPGISpech** | **Tedlium** | **Voxpopuli** | ... | `Canary-1b-v2` | 749 | 7.15 | 16.01 | 10.82 | 2.18 | 3.56 | 11.79 | 2.28 | 4.29 | 6.25 |” (Open ASR Leaderboard table, LibriSpeech test-other column)
- NVIDIA Canary 1B v2 had an average word error rate of 7.27 percent on the Multilingual LibriSpeech test across its MLS-6 languages.
“| **WER ↓** | Fleurs-25 Langs | CoVoST-13 Langs | MLS - 6 Langs | ... | **`Canary-1b-v2`** | 8.40% | 8.85% | 7.27% |” (ASR table, MLS-6 column)
- NVIDIA Canary 1B v2 had a word error rate of 5.08 percent on LibriSpeech Clean with MUSAN noise at 0 dB SNR.
“Performance across different Signal-to-Noise Ratios (SNR) using MUSAN music and noise samples \[16] on the [LibriSpeech Clean test set](https://www.openslr.org/12). ... | **SNR (dB)** | 100 | 10 | 5 | 0 | -5 | ... | **`Canary-1b-v2`** | 2.18% | 2.29% | 2.80% | 5.08% | 19.38% |” (Noise robustness table, SNR 0 dB column)
- NVIDIA Canary 1B v2 had a word error rate of 2.29 percent on LibriSpeech Clean with MUSAN noise at 10 dB SNR.
“Performance across different Signal-to-Noise Ratios (SNR) using MUSAN music and noise samples \[16] on the [LibriSpeech Clean test set](https://www.openslr.org/12). ... | **SNR (dB)** | 100 | 10 | 5 | 0 | -5 | ... | **`Canary-1b-v2`** | 2.18% | 2.29% | 2.80% | 5.08% | 19.38% |” (Noise robustness table, SNR 10 dB column)
- NVIDIA Canary 1B v2 had a word error rate of 2.80 percent on LibriSpeech Clean with MUSAN noise at 5 dB SNR.
“Performance across different Signal-to-Noise Ratios (SNR) using MUSAN music and noise samples \[16] on the [LibriSpeech Clean test set](https://www.openslr.org/12). ... | **SNR (dB)** | 100 | 10 | 5 | 0 | -5 | ... | **`Canary-1b-v2`** | 2.18% | 2.29% | 2.80% | 5.08% | 19.38% |” (Noise robustness table, SNR 5 dB column)
- NVIDIA Canary 1B v2 had a word error rate of 19.38 percent on LibriSpeech Clean with MUSAN noise at -5 dB SNR.
“Performance across different Signal-to-Noise Ratios (SNR) using MUSAN music and noise samples \[16] on the [LibriSpeech Clean test set](https://www.openslr.org/12). ... | **SNR (dB)** | 100 | 10 | 5 | 0 | -5 | ... | **`Canary-1b-v2`** | 2.18% | 2.29% | 2.80% | 5.08% | 19.38% |” (Noise robustness table, SNR -5 dB column)
- NVIDIA Canary 1B v2 had a word error rate of 2.28 percent on the SPGI Speech test set in the Open ASR Leaderboard table.
“| **WER ↓** | **RTFx** | **Mean** | **AMI** | **GigaSpeech** | **LS Clean** | **LS Other** | **Earnings22** | **SPGISpech** | **Tedlium** | **Voxpopuli** | ... | `Canary-1b-v2` | 749 | 7.15 | 16.01 | 10.82 | 2.18 | 3.56 | 11.79 | 2.28 | 4.29 | 6.25 |” (Open ASR Leaderboard table, SPGI Speech column)
- NVIDIA Canary 1B v2 had a word error rate of 4.29 percent on the TED-LIUM test set in the Open ASR Leaderboard table.
“| **WER ↓** | **RTFx** | **Mean** | **AMI** | **GigaSpeech** | **LS Clean** | **LS Other** | **Earnings22** | **SPGISpech** | **Tedlium** | **Voxpopuli** | ... | `Canary-1b-v2` | 749 | 7.15 | 16.01 | 10.82 | 2.18 | 3.56 | 11.79 | 2.28 | 4.29 | 6.25 |” (Open ASR Leaderboard table, TED-LIUM column)
- NVIDIA Canary 1B v2 had a word error rate of 9.87 percent on the This American Life long-form recordings.
“| **Dataset** | **WER ↓** | ... | This American Life | 9.87% |” (Long-form inference table, This American Life row)
- NVIDIA Canary 1B v2 had a word error rate of 6.25 percent on the VoxPopuli test set in the Open ASR Leaderboard table.
“| **WER ↓** | **RTFx** | **Mean** | **AMI** | **GigaSpeech** | **LS Clean** | **LS Other** | **Earnings22** | **SPGISpech** | **Tedlium** | **Voxpopuli** | ... | `Canary-1b-v2` | 749 | 7.15 | 16.01 | 10.82 | 2.18 | 3.56 | 11.79 | 2.28 | 4.29 | 6.25 |” (Open ASR Leaderboard table, VoxPopuli column)