NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent, Promptable Multilingual Transcription
NVIDIA's Nemotron 3.5 Transcribe family, paired with the Granary v2 dataset, delivers promptable, adaptable multilingual speech recognition. Together they show how far automatic speech recognition (ASR) has moved by 2026: from fixed, one-size-fits-all transcription toward systems that respond to instructions and adjust to domain-specific vocabulary and context.
The Model at a Glance
- Repository:
nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B - Task: Automatic Speech Recognition (ASR)
- Parameter count: 34B reported on the model card
- Update: Updated October 28, 2026
- Format: Hosted on Hugging Face, distributed alongside companion data resources
Why It Matters
Traditional ASR pipelines transcribe audio in a largely fixed way and require separate components—custom vocabularies, diarization tools, and post-processing—to handle real-world nuance. Nemotron 3.5 Transcribe takes a more unified approach: the model can be prompted, so users can steer output style, terminology, formatting, and language handling through natural-language instructions.
That matters for 2026 deployments where transcription feeds directly into downstream systems—meeting summaries, compliance archives, accessibility captions, and voice agents—and where accuracy on names, jargon, and mixed-language speech is often the deciding factor.
Promptability and Adaptability
- Instruction-driven output: Specify formatting, casing, punctuation, and style directly in the prompt.
- Vocabulary steering: Bias recognition toward domain terms, product names, and proper nouns without retraining.
- Multilingual coverage: Handle multilingual and code-switched audio within a single adaptable model.
- Rapid reconfiguration: Adapt to a new domain or locale through prompting rather than a full fine-tuning cycle.
Granary v2: The Data Foundation
Granary v2 is NVIDIA's companion data resource supporting the model. Large, carefully curated and labeled audio corpora are the foundation of reliable ASR, and Granary v2 is positioned to support broad multilingual coverage and the prompt- and context-conditioned behavior that defines this generation of models. Paired with the model release, it underscores the shift toward data-and-model bundles that make adaptation practical for downstream teams.
Practical Implications
For teams building in 2026, the combination points to a more flexible transcription stack:
- Less custom plumbing: Fewer auxiliary tools for vocabulary and formatting.
- Faster domain onboarding: Prompt changes instead of retraining for new terminology.
- Better multilingual reach: One model serving multiple languages and code-switched content.
- Cleaner downstream integration: Output shaped for the systems that consume it.
Looking Ahead
Nemotron 3.5 Transcribe and Granary v2 reflect a broader trend: speech models are becoming controllable interfaces rather than fixed black boxes. As prompting and data quality continue to improve, expect transcription systems that adapt to users, domains, and languages with far less engineering overhead.
See the model card at nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B on Hugging Face for download details and usage guidance.