Fact · Speech recognition
19.38percent word error rate
NVIDIA Canary 1B v2 had a word error rate of 19.38 percent on LibriSpeech Clean with MUSAN noise at -5 dB SNR.
Quote found word for word on the live source, last checked October 19, 2026
In the source's own words
“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% |”
canary-1b-v2 model card (raw README), NVIDIA, Hugging Face (NVIDIA model card), 2025, Noise robustness table, SNR -5 dB column.
Open the original- What it measures
- The card adds MUSAN music and noise to the LibriSpeech Clean test set at the stated signal-to-noise ratio (SNR) and reports the WER for each level.
- What it does not show
- Noise is added artificially to clean recordings, so it is not a recording made in a noisy room.
In the data
A row of Speech Recognition Accuracy: Word Error Rates, as Canary 1B v2, noise -5 dB: 19.38 percent.
- System
- NVIDIA Canary 1B v2
- Kind
- Open model
- Test set
- LibriSpeech test-clean with MUSAN noise
- Language
- English
- Speech
- Noisy or distant
- Year
- 2025