NVIDIA opens Nemotron 3.5 Transcribe: promptable open-weight multilingual ASR with diarization, translation, and audio Q&A; Granary v2 data pipeline

Why it matters

【Key points】

On 2026-10-28 NVIDIA published Nemotron 3.5 Transcribe: an open-weight, promptable multilingual speech model unifying transcription, speaker attribution, translation, and audio Q&A. Granary v2 data pipeline and NeMo Speech fine-tune recipes ship with it. Recorded audio (not live streaming); BF16/FP8/NVFP4; runs on RTX/DGX Spark and Jetson AGX Thor.

【Verifiable】

- NVIDIA blog: https://blogs.nvidia.com/blog/local-ai-rtx-spark-laptops-updates-models-agents-games/ (RSS 2026-10-28T15:00:46Z)

- HF post: https://huggingface.co/blog/nvidia/nemotron-transcribe-granary/ (Oct 28, 2026)

- Model card: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B

- Vendor/leaderboard cites (not re-run here): ~40 languages ASR; ~31 EN translation; Sortformer diarization. HF Open ASR (EN, open-weight) ~5.33% avg WER; Artificial Analysis STT ~2.8% AA-WER at ~4B active; David AI 21-lang DER ~18.26%; VoiceArena cpWER lead. Medical case ~45.8% relative WER cut vs best prompt.

- Peers on site: e/1051 Bolt E4B; e/776 MAI-Transcribe.

【Our read】

Speech stacks are shifting from ASR plus post-hoc diarize/MT toward single-checkpoint natural-language control. Whisper-class, MAI-Transcribe, and Voxtral 3 already compete; NVIDIA adds an open, fine-tunable, promptable line tied to RTX Spark.

Peers: general ASR is strong on words but weak on unified diarize+MT+QA; MAI-Transcribe leans streaming/product; Voxtral sits in Mistral’s audio family. Open-board numbers look strong but are vendor-cited slices—don’t merge into “#1 everywhere.” ~4B active params favors edge/Spark.

Impact: fewer glued models for multilingual meeting/medical/classroom pipelines; pairs with local agents. Granary v2 extends control into training. Limits: not streaming; quantization/tooling still rolling; demos may diverge from independent verify. This item is Transcribe+Granary only, not laptop marketing.