Canary-1B-v2

Canary-1B-v2 is NVIDIA's billion-parameter multitask speech model for high-quality transcription and translation across 25 European languages. Trained on the Granary dataset of around a million hours of audio built with Carnegie Mellon University and Fondazione Bruno Kessler, it handles automatic speech recognition plus speech translation in both directions between English and the 24 other supported languages.

The model is released under CC-BY-4.0, which permits commercial use, and NVIDIA reports English transcription accuracy competitive with Whisper-large-v3 at a fraction of the size. It has become a popular open alternative for production ASR pipelines and multilingual voice agents.