Nvidia released Nemotron 3.5 Transcribe 30B-A4B on 28 October 2026 — an open-weight model turning speech into text in 40 languages, Polish among them. It handles recordings up to an hour long and works out who said what, but deliberately does not run on live streams. A second version of the Granary training dataset shipped alongside it.
Key takeaways
- 34B parameters in total, roughly 4.3B active per forward pass — audio encoder included
- 40 transcription languages, speech translation both ways with 31 languages
- Recordings up to 1 hour, a 65,536-token context window, one H100 80 GB card
- Mean WER of 4.71% on the English Open ASR Leaderboard against 4.94% for Qwen3-ASR-1.7B
- OpenMDW-1.1 licence, weights in BF16, FP8 and NVFP4
Picture the following situation
A company records an hour-long call with four participants — an editorial example, not a case from the sources. It goes to the model with one line of instruction: produce a clean transcript without fillers, with punctuation, speaker labels and a timestamp on every word.
The model returns finished text in one pass, replacing three tools — speech recognition, speaker separation and text cleanup.
But a problem appears: the same company would also like captions during the meeting, and Nemotron 3.5 Transcribe will not do that.
The model card is blunt: this is an offline model, and splitting audio internally into 30-second chunks does not turn it into a streaming service.
What exactly did Nvidia release?
Nemotron 3.5 Transcribe 30B-A4B takes a recording plus a text instruction and returns text. Nvidia classifies it as a Speech-Augmented Language Model — a language model with a listening module bolted on. The difference from classic speech recognition: output style is set by a prompt sentence, not a code switch.
Three components sit underneath. An ASR encoder of roughly 600M parameters converts audio into a representation the language model understands. Beside it, the Nemotron-3-Diarization speaker encoder (roughly 99M parameters) tracks who speaks. Both feed the backbone?backbone: the main language model the whole system is built on Nemotron 3.5 Lightning 30B-A3B — a Mamba-2 and transformer hybrid in a mixture-of-experts layout.
Of 34B total parameters, roughly 4.3B are active on any single pass: about 3.5B in the backbone plus about 800M in the audio encoder. Hence the model fits on one H100 80 GB card, and the documentation also lists sub-server hardware — Jetson Thor, DGX Spark (GB10) and the RTX 5090 at NVFP4.
The diagram shows the flow described in the model card, not a specific deployment.
Does the quality actually beat the competition?
WER (word error rate) is the share of words recognised incorrectly — lower is better. On the English Open ASR Leaderboard, a set of nine public corpora, Nemotron records a mean 4.71% error rate against 4.94% for Qwen3-ASR-1.7B, the best open-weight model so far. A real lead, but narrow.
The gap widens on long or non-English recordings. On long-form material (CORAAL, Earnings21, Earnings22) the mean WER is 10.98% against 13.37% for Parakeet-TDT-0.6B-v3, and on the multilingual leaderboard 4.62% against 6.49% for Qwen3-ASR-1.7B.
The most interesting result covers a task a single model usually does not perform. With several speakers what counts is cpWER — WER computed only after speaker labels are matched to real people. The mean across four corpora (AMI, CALLHOME, Mixer 6, NOTSOFAR-1) is 15.80%, the best published result, ahead of the DiCoW v3.3 and DiariZen pair at 16.83%.
| Benchmark | Nemotron 3.5 Transcribe | Best previous result |
|---|---|---|
| Open ASR Leaderboard (English) | 4.71% WER | 4.94% — Qwen3-ASR-1.7B |
| Multilingual leaderboard | 4.62% WER | 6.49% — Qwen3-ASR-1.7B |
| Long-form material (CORAAL, Earnings21, Earnings22) | 10.98% WER | 13.37% — Parakeet-TDT-0.6B-v3 |
| Multi-speaker (AMI, CALLHOME, Mixer 6, NOTSOFAR-1) | 15.80% cpWER | 16.83% — DiCoW v3.3 and DiariZen |
What about Polish?
Polish landed in the highest of the model card's three quality tiers — "Leading", meaning close to state of the art without fine-tuning. On FLEURS?FLEURS: a multilingual speech test set used to compare speech recognition models it records 3.81% WER for Polish and identifies the language correctly in 99.87% of cases.
What is missing: speaker attribution is confirmed for thirteen languages, Polish not among them. Speech translation both ways with English already works — on FLEURS, Polish into English gives 26.54 BLEU?BLEU: a machine translation quality score, higher meaning closer to a reference translation, the reverse 18.56.
What do these names mean?
- MoE (mixture of experts) — a model split into specialised blocks, only a few firing per token. Like a large firm where two specialists handle one ticket, not the whole team.
- Diarization — working out who speaks in which part of a recording. Not identity recognition: the model assigns
<spk:0>,<spk:1>tags without knowing who they belong to. - CTC head — an extra module attaching word-level timestamps to the transcript.
- NVFP4 — Nvidia's four-bit number format. Weights take less memory at the cost of some precision, letting the model fit on a smaller card.
- Speculative decoding — a smaller model guesses several tokens at once and the large one only verifies them. Cuts response time without changing output.
All of these serve one goal: fitting 34-billion-parameter quality into a few-billion-parameter budget.
Why isn't a plain speech recognition model enough?
Chaining blocks
A meeting transcript today is usually assembled from several blocks: one model turns audio into words, a second separates speakers, a third sorts punctuation and numbers. The proof is Nvidia's own table — the best previous multi-speaker result belongs not to one model but to the DiCoW v3.3 and DiariZen pair.
Chaining blocks has a cost: every boundary is a place where one stage's error feeds the next.
Vocabulary
A model trained on general speech mangles drug names, product codes and identifiers. The classic workaround is a decoder word list or fine-tuning. Here the terms go into the prompt, which in Nvidia's earnings-call tests lifts keyword recall from 58.16% to 96.05%.
Style
A verbatim transcript helps in research, but captions need clean text — normally a separate stage. Nemotron switches modes by instruction, and the Preference-ASR benchmark measures it directly: 7.56% WER without an instruction, 6.16% with one.
Is it working yet?
Yes — the weights have been publicly available since 28 October 2026. Nvidia published the model on Hugging Face, NGC, build.nvidia.com and the NeMo Speech repository under the OpenMDW-1.1 licence, which permits commercial use. Alongside BF16 there are FP8 and NVFP4 variants plus a DFlash2 helper for speculative decoding.
One caveat concerns throughput. Nvidia states that on an RTX PRO 6000 Blackwell Server Edition card the model processes up to 2,238 seconds of audio per second and returns a 30-second transcript in 0.13 s — but those serving optimisations did not reach the open source release. The shipped vLLM integration is meant for development and, per the model card, runs roughly twice as slowly. Second caveat: the missing live mode is a design property, not a temporary limitation.
Why does this matter?
Meeting transcription and medical documentation are two areas where AI already earns money, and both so far required either the cloud or a pipeline stitched from several models. Open weights fitting on a single card push that choice toward local deployment — which matters wherever recordings cannot leave the organisation. The missing live mode leaves a whole category of voice assistants out of reach.
What's next?
- Nvidia states in the model card it intends to upstream the optimised serving path into NeMo Speech and vLLM — no date given
- The Granary v2 pipeline is already public in NeMo Curator, with end-to-end runs and per-stage scripts to follow later
- Multi-speaker transcription for Polish and twenty-plus other languages remains a gap to close by fine-tuning in NeMo Speech
Sources
- TechCrunch — Voice AI in vogue: Nvidia releases new open transcription model
- Hugging Face — NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
- Hugging Face — Granary-v2
