One checkpoint, four jobs, shrinking language counts
Most transcription pipelines are chains: one tool writes the words, a second works out who said them, a third translates, and a summary is a fourth step with its own errors. Transcribe is built to collapse that chain. A prompt can ask for a verbatim transcript with speaker labels, a translation into English, the decisions taken in a meeting, or a transcript that spells supplied terms, such as drug names, the way a clinic writes them. That promptability, rather than any single score, is what separates this release from Nvidia's smaller Parakeet line and from a conventional speech-to-text model.
The language headline needs its table. Forty languages is the transcription count. Translation covers 31 of them, to or from English. Speaker attribution, the feature meeting software actually sells, covers 13. Audio question answering was trained in English. None of that makes the model weak; it makes the one-number version of the release wrong, and anyone planning a deployment in, say, Cantonese or Hindi should plan from the task-by-language matrix in the model card rather than from the announcement headline.
The speed number belongs to a version you cannot download yet
The figure most likely to travel is throughput: 100 hours of audio processed in under three minutes on a single data-centre GPU. Nvidia's announcement attaches two conditions to it in the same paragraph. The run used the forthcoming NVFP4 quantized version of the model, and it was measured on 30-second clips at high concurrency. Quantized versions are listed as coming soon, and local deployment tooling is coming with them. The checkpoint available today is the full-precision one, servable through vLLM, and the model card notes that the serving optimizations behind the fastest throughput numbers are not yet part of the open-source release.
The throughput figure is a disclosed benchmark result for a configuration that is not yet available in the released checkpoint. It should not be treated as measured performance of the version users can download today. It may well arrive as promised. The mixture-of-experts architecture activates roughly 4B parameters per token, according to Nvidia, despite the much larger total checkpoint. That may help explain its throughput potential, but active parameter count alone does not establish end-to-end inference speed or cost. But a buyer comparing transcription costs this week should price the version that exists, at the clip lengths they actually have. Hour-long recordings are supported; they are not what the speed figure was measured on.
Where the scores are strong, and where they thin out
Nvidia reports strong results across several third-party evaluations. It says Nemotron 3.5 Transcribe ranks first among open-weight models on the Hugging Face English Open ASR Leaderboard at 5.33% average WER, with API models excluded. On Artificial Analysis's AA-WER evaluation, Nvidia reports a tie for first among open-weight models at 2.8%, alongside Mistral's Voxtral Small. Nvidia lists approximately 4B active parameters per token for Transcribe, compared with 24B for Voxtral Small; that is a model-architecture comparison, not a direct measurement of serving efficiency. The long-form track is the sobering one: Nvidia reports 10.98% mean word error rate there, and long-form audio is what meetings and dictation actually are. Nvidia also reports first place on VoiceArena's English diarization benchmark, at a 32.4% combined error rate once speaker labels are counted, which says as much about the state of diarization as it does about this model.
Nvidia's medical case study used an Eka Care benchmark containing 3,619 clips. On the full set, word error rate fell from 30.8% without prompting to 23.8% with a language-and-alphabet prompt, and to 11.8% after fine-tuning. But that full-set result includes the 855 clips used for adaptation. On the remaining 2,764 held-out clips, the best-prompt baseline was 22.44% WER, falling to 12.17% after fine-tuning—a 45.8% relative reduction. The held-out result is the more meaningful generalization test, although the announcement does not establish a speaker-disjoint evaluation. Nvidia reports approximately 14 GPU-hours of fine-tuning on one GB300, using about 153 hours of synthetic audio, 29 hours of accented English and 63 hours of medical audio. Those figures describe the reported adaptation workload, not the total cost of collecting, preparing, validating and deploying a medical transcription system. The study also records the costs of adapting: general English error rose slightly after fine-tuning, and a specialist medical model still beats the adapted model on brand-name drugs. That is what a real evaluation looks like, and it is the standard the rest of the launch figures should be read against.
Sources
- [1] Nvidia, "NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt," Hugging Face, October 28, 2026 (primary source; all capability counts, leaderboard figures, speed conditions, Granary v2 figures and the medical case study are from this announcement)Read source
- [2] Unite.AI, "NVIDIA Launches Open-Weight Speech Model and 463,062-Hour Audio Dataset," October 28, 2026 (model-card details: 34B full checkpoint, OpenMDW-1.1 licence, training scale, serving-optimization caveat)Read source
- [3] TechCrunch, "Voice AI in vogue: Nvidia releases new open transcription model," October 28, 2026 (independent report; release timing, recorded-audio limitation, voice-model context)Read source