Nemotron 3.5 Transcribe: open ASR lasts only if Granary-style curation ships with it

On 28 October 2026 NVIDIA published Nemotron 3.5 Transcribe on the Hugging Face blog: an open-weight, promptable speech model that folds transcription, speaker attribution, translation and audio Q&A into one checkpoint—and ships Granary v2 multilingual data with a reusable curation pipeline alongside it.

Leaderboard wins are the headline. The more durable claim is that domain gaps do not close by shaving another WER point alone. Format targets, curation metadata and fine-tune recipes must be open so prompt-level control extends into training-level control.

Capability matrix: ASR in 40 languages; translation for 31 of them to or from English; diarization in 13; recordings up to about an hour with up to eight speakers; word-level timestamps. The same checkpoint switches, by instruction, among labeled meeting notes, English translation, or “what decisions were made.”

Architecturally, a speech encoder pairs with a hybrid Mamba–Transformer MoE language model so roughly 4B of 30B parameters are active per token. The model targets recorded audio, not live streaming; quantized builds and NeMo-Speech.cpp local deploy are forthcoming, while today’s checkpoint serves via vLLM and similar stacks.

English Open ASR open-weight lead at 5.33% average WER; model-card Open ASR comparisons also cite lowest open-weight multilingual (4.62%) and long-form (10.98%) means. Artificial Analysis: tied open-weight first at 2.8% AA-WER, 1.8% on AA-AgentTalk. David AI 21-language diarization: lowest DER among 14 systems at 18.26%; VoiceArena 14-system lead at 32.4% cpWER with deletions at 8.32% of reference words (median ~31%).

Granary v2’s public manifest spans 463,062 hours across 48 languages and 16 source corpora, emitting three transcript forms per clip (verbatim; ITN with disfluencies; ITN clean) plus captions, code-switch restoration, speech-grounded Q&A and quality-scored translations. Medical case study on Eka Care’s Indian clinical benchmark: best-prompt held-out WER 22.44% → 12.17% after fine-tune (−45.8% relative); full adaptation ~14 GPU-hours on one GB300.

Steelman: leaderboard figures arrive via a vendor-friendly post; some comparisons exclude API models or rely on model-card tables; Granary numbers describe the dataset release, not Transcribe’s full training mixture; medical gains use synthetic speech on one benchmark and may not travel.

Still thin: full mixture weights; error decomposition vs dedicated diarizers on production overlap; post-quantization WER retention; third-party replication of the prompt→fine-tune curve outside Indian medical speech. Confidence in the posted HF/AA/David numbers is high; confidence that “open the pipeline and domains close” is medium.

For enterprise ASR buyers, one promptable checkpoint cuts multi-model orchestration cost, but domain vocab still needs real labeled clips and a checkpoint-selection set (the post suggests ≥100). For open ecosystems, Granary-style transcript forms, quality flags and task labels are closer to auditable training data than raw dumps. For closed speech APIs, open-weight placement on the accuracy–active-parameter Pareto keeps pressure on price and private-deploy narratives.

Watch whether quantization and local runtimes keep Open ASR / AA rank; whether the community reproduces similar relative WER cuts in legal, support and dialect domains with Granary; and whether “promptable alone” reduces dependence on cascaded diarization in real overlapping meetings.