Canary (NVIDIA NeMo)

Canary is NVIDIA's family of multilingual speech models built on the NeMo toolkit, combining a FastConformer encoder with a Transformer decoder in an encoder-decoder design. The original Canary-1B has one billion parameters across 24 encoder and 24 decoder layers and covers English, German, French, and Spanish, performing both transcription and speech-to-text translation between English and the other three languages, with or without punctuation and capitalization. Training drew on 85,000 hours of speech, and the model uses concatenated SentencePiece tokenizers, one per language. Later releases extended the family to more languages with lower word error rates. Models load through NeMo, an open source Apache 2.0 speech framework, with tasks and languages specified through prompts or manifest files; weights are published on Hugging Face under their own terms, and the original 1B checkpoint carries a CC-BY-NC-4.0 license. Speech teams reach for Canary when transcription and translation are needed from a single model.