NVIDIA released Nemotron 3.5 Transcribe alongside Granary v2, a paired dataset designed to train and adapt multilingual speech recognition systems you can steer with natural language prompts. The model supports over 40 languages and is built on a Canary-1B architecture fine-tuned on Granary v2, which contains roughly 75,000 hours of curated, filtered multilingual audio.
The prompt-steering capability is the technical detail worth examining closely. Instead of fixed decoding behavior, users can instruct the model to output punctuated text, adjust formality, or handle domain-specific vocabulary, all at inference time without retraining. Granary v2 itself is documented in enough detail to show exactly how NVIDIA filtered and weighted source audio, which matters if you are building your own training pipeline and need a reproducible baseline.
The full post walks through benchmark comparisons against Whisper large-v3 and SeamlessM4T across multiple language families, showing where Nemotron 3.5 wins and where it does not. Those numbers, combined with the dataset construction methodology, are the reason to read past the headline.
[READ ORIGINAL →]