Term
NVIDIA Canary
Canary is NVIDIAs open ASR and speech-translation model family (CC-BY-4.0) that tops the Open ASR Leaderboard on English accuracy ahead of OpenAI Whisper and covers 25 European languages.
NVIDIA Canary — explained in more detail
Canary is an open model family from NVIDIA for automatic speech recognition (ASR) and speech-to-text translation (AST). Unlike closed cloud services, the weights are publicly available on Hugging Face under the permissive CC-BY-4.0 license and may be used commercially and non-commercially. The models use a FastConformer encoder with a Transformer decoder and run in the NeMo framework. The Canary-1B-v2 variant released in August 2025 has around 978 million parameters and supports 25 European languages.
Notable is its accuracy relative to training effort: Canary outperforms Whisper, OWSM and Seamless-M4T on English, French, Spanish and German, despite being trained on an order of magnitude less data. On the Open ASR Leaderboard the family leads on English accuracy ahead of OpenAI Whisper (large-v3). This makes Canary a strong open alternative that enables local operation and full control.
Example / In practice
A company with data-protection requirements can run Canary locally or in its own cloud instead of sending audio to an external service. This allows, for example, transcribing internal meetings or multilingual recordings on premise and translating between the 25 supported languages, without per-minute API costs.
Distinction from similar terms
Canary stands as an open-weight family in contrast to the closed services AssemblyAI Universal-2 (accuracy leader, managed service) and Deepgram Nova-3 (real-time streaming leader). Compared with OpenAI Whisper, Canary is stronger on English accuracy but covers fewer languages with 25 versus Whispers roughly 99. The companion Parakeet family targets even smaller, faster models.