NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt
- ID
- 40394
- Status
- summarized
- Published
- 28 Oct 2026, 10:58 PM
- Fetched
- 28 Oct 2026, 11:58 PM
- Provider
- Hugging Face Blog
- Category
- developer-ai
- Original URL
- https://huggingface.co/blog/nvidia/nemotron-transcribe-granary
- Source URL
- https://huggingface.co/blog/feed.xml
Summary
- Score
- 6.5
- Created
- 28 Oct 2026, 11:58 PM
- Tags
- Audience
- developersai_ml_learnerssaas_founders
What happened
NVIDIA released Nemotron 3.5 Transcribe as an open-weight, promptable speech model on Hugging Face and ModelScope, combining transcription (40 languages), translation to/from English (31 languages), speaker diarization (13 languages, up to 8 speakers), and audio question answering in one checkpoint, handling recordings up to an hour on a single NVIDIA GPU. NVIDIA claims first place among open-weight models on the Hugging Face Open ASR Leaderboard for English and top spots on David AI's 21-language diarization leaderboard and VoiceArena's 14-system comparison. Alongside it, Granary v2's multilingual training data and curation pipeline are released, and a medical fine-tuning case study reports a 45.8% relative word error rate cut on held-out speakers versus the best prompt.
Why it matters
If you currently pay per-minute for hosted ASR plus a separate diarization or translation vendor, this is one open checkpoint you can run on your own GPU, which also means audio never leaves your infrastructure — a concrete answer to client data-residency objections. The 45.8% relative WER drop from fine-tuning versus prompt-only is the real decision point: budget fine-tuning on your own domain audio rather than assuming prompting is enough. Note what the post does not say: the 40 languages aren't enumerated, so verify Malay, Manglish and code-switched call audio yourself before committing.
Discussion angle
Run the same 10 minutes of real Malaysian call-centre or meeting audio through this and through your current hosted ASR: does the promptable diarization actually hold up on code-switched Malay/English, and at what GPU cost per hour versus per-minute API pricing?