Nemotron 3.5 Transcribe: NVIDIA's Speech LLM Beats the Diarization Cascade — in 13 Languages

NVIDIA shipped Nemotron 3.5 Transcribe on October 28, 2026 under an open licence, and the number in every headline about it is 40: forty languages, one checkpoint, speaker labels included. The number that actually changes how you build a transcription stack is 15.80. That is the model's mean cpWER on NVIDIA's own multi-speaker benchmark set, against 16.83 for the best specialised cascade the company compared it with — a speech recogniser plus a separate diarisation pipeline, wired together the way most production meeting-transcription systems still are. NVIDIA claims this is the first speech LLM to beat that arrangement, and it is the claim worth testing, because it is the one that removes a component from your architecture rather than adding a language to a list.

The 40-language figure is real but narrower than it reads. Speaker attribution — the part that replaces the cascade — is trained for thirteen languages, not forty. The model reads recorded files only, up to an hour at a time, with no streaming path. And it is not a small download: 34B parameters in total, about 4.3B active per token, with an H100 80GB as the stated baseline for BF16 serving. Those three constraints decide whether the diarisation result is something you can use this quarter or something you note for the next one.

What NVIDIA actually shipped

The checkpoint is published as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, and the gap between that name and the published parameter count is the first thing to clear up. The model is built on Nemotron 3.5 Lightning 30B-A3B, a hybrid Mamba-2 and Transformer mixture-of-experts language model; NVIDIA's model card states a total of 34B parameters, with roughly 4.3B active including the audio stack. The extra weight over the 30B in the name is the encoders: about 600M for the ASR transformer and about 99M for the speaker encoder, roughly 800M together, leaving the language backbone at about 3.5B active. Several write-ups have reported this as an inconsistency; it is not one, it is a name that counts the LLM and a card that counts the whole system.

Architecturally, the piece worth understanding is the parallel expert encoder. A frozen Nemotron 3 Diarization model produces speaker-activity representations, and those are injected into the ASR encoder's output, so the language model sees both the words and who is likely speaking them in a single pass. A lightweight CTC adapter then aligns the generated text back to the audio to produce speaker-attributed word timestamps. There is no second alignment stage and no pyannote call in the middle. Native multi-token prediction doubles as a speculative-decoding draft inside vLLM, and NVIDIA also ships a separately trained DFlash2 draft model.

• Total and active parameters — 34B total, ~4.3B active per token, of which ~3.5B is the language backbone

• Context and audio — 65,536 tokens, recordings up to 1 hour at 16 kHz mono, decoded in 30-second internal chunks

• Language coverage — 40 languages for transcription, 32 for translation to and from English, 13 for speaker-attributed transcription

• Licence — OpenMDW-1.1, commercial and non-commercial use permitted

• Serving baseline — one H100 80GB at BF16; BF16, FP8 and NVFP4 builds are published

• Streaming — none; this is a file-in, transcript-out model

The release is a pair of artefacts, and it is worth keeping them apart because most coverage blurs them. Alongside the model, NVIDIA published Granary v2, a speech-data pipeline release rather than a checkpoint: 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1, built with NeMo Curator and shipped with a public manifest. Each record carries three transcript forms — text-normalised verbatim, inverse-normalised with disfluencies, and inverse-normalised without them — plus task labels, speech-grounded question-answer pairs, contextual biasing terms and a per-record quality flag. It is the data the fine-tuning recipes expect, not weights you can run. The source audio still has to be downloaded under each corpus's own licence, so Granary v2 is open in the sense that its manifest and pipeline are, not in the sense that the audio is free to redistribute.

The licence matters as much as the weights. OpenMDW-1.1 permits commercial use, which puts this in the same bracket as Mistral's Voxtral line rather than the research-only end of the open-weight spectrum. Runtime support at launch is broader than most speech LLM releases: vLLM 0.32.0 is the primary path and the only one that gives you the full feature set, with word timestamps requiring the V1 runner and timestamp middleware and no DFlash2. SGLang-Omni 0.2.0 on SGLang 0.5.22 handles whole-file transcription through the OpenAI-compatible audio endpoint but not word timestamps, speaker-attributed output, translation or speculative decoding. NeMo Speech ships the fine-tuning toolkit, Transformers supports in-process use from a source install, and a GGML build covers selected Blackwell parts including the RTX 5090 and RTX PRO 6000.

The diarisation result is the story

NVIDIA's multi-speaker table is where the cascade comparison lives, and it is unusually detailed for a launch post. Mean cpWER across the set is 15.80 against 16.83 for DiCoW v3.3 combined with DiariZen. The per-condition numbers show where the margin comes from and where it does not: AMI in close-talk conditions reaches 10.29, AMI single-distance-microphone 17.34, CALLHOME CH109 11.50, Mixer 6 channel 4 14.80, NOTSOFAR-1 close-talk 14.94 and single-channel 25.91, with AISHELL-4 at 32.91 cpCER and AliMeeting at 34.95. The wideband, multi-channel meeting case is strong; far-field single-microphone and Chinese meeting audio are much weaker, which is a fair description of where every diarisation system struggles.

The independent-looking corroboration NVIDIA leans on is VoiceArena's Diarization-Bench, where the model reports 32.4% cpWER across English full sessions, first of fourteen systems. The runner-up, Multitalker Parakeet paired with Nemotron 3 Diarization, is at 34.2%, and the rest of the field runs from 34.7% to 71.9%. NVIDIA also reports word deletions at 8.32% of reference words against a field median around 31%. That second figure is the more useful one for anyone who has shipped a transcript: cpWER can be gamed by dropping words you are unsure about, and a system that keeps 92% of the reference words while winning the metric is doing the harder thing. Both are vendor-reported numbers on a third-party benchmark, not an independent evaluation of the model.

The limit is the language count. Speaker attribution is trained for thirteen languages — English, French, German, Hindi, Italian, Korean, Portuguese, Russian, Spanish, Japanese, Thai, Arabic and Chinese — and supports up to eight speakers. If your recordings are English meetings or one of those twelve others, the cascade-replacement argument is live. If your audio is Polish, Turkish or Vietnamese, you get transcription without diarisation and you still need the second component.

Where it actually ranks: three leaderboards, three different answers

The English number depends on which NVIDIA page you read, and it is worth being precise about that rather than picking the better one. The model card's nine-set mean word error rate is 4.71%, with Qwen3-ASR-1.7B next at 4.94%. NVIDIA's Hugging Face write-up from the same day quotes 5.33% as the average WER on the Hugging Face Open ASR Leaderboard, at an RTFx of 462.09, with Qwen3-ASR-1.7B at 5.4%. Both are the vendor's own figures, published within hours of each other, and the difference is which test sets are in the average. Do not average them, do not subtract one from the other, and do not treat the decimal as settled. What survives both framings is the placement: first among open-weight models on that board.

The rest of the vendor-reported set is consistent in shape. Multilingual mean WER is 4.62% across seven languages against 6.49% for Qwen3-ASR-1.7B. Long-form mean is 10.98% against 13.37% for Parakeet-TDT-0.6B-v3. Style control scores 7.56% default and 6.16% when instructed, against Canary-Qwen at 8.16% and 8.36%. Contextual prompting moves ContextASR-Bench from 3.33% with no context to 1.71% with fine-grained context, and earnings-call transcription from 12.54% to 9.58% with local context supplied.

The independent check tells a less flattering story, and it is the one to read if you are choosing between models this week. On Artificial Analysis's non-streaming word-error index, checked on October 29, Nemotron 3.5 Transcribe reads 2.8%. Voxtral Small Transcribe 3 reads 2.1% and MAI-Transcribe-2 reads 2.0% on the same view, while Gemini 3.5 Transcribe sits at 2.6%. NVIDIA's "tied for first among open-weight models" framing is defensible against the specific model it names — Voxtral Small, which the same index also shows at 2.8%, and which NVIDIA describes as a 24B model against this one's roughly 4B active parameters — but the leaderboard moved past that pairing before this model shipped. A newer Mistral sibling, Voxtral Small Transcribe 3, is about a quarter lower in error rate on the same view. That does not make the diarisation result wrong, because Voxtral is not doing speaker attribution in one pass. It does mean the right sentence is "best open-weight model on speaker-attributed ASR" and not "best open-weight transcriber."

The per-language tail is where deployments break

Coverage and parity are different claims, and NVIDIA's model card publishes enough per-language detail to see the distance between them. The language matrix has 41 rows; Norwegian appears in it without task checkmarks, leaving 40 languages actually marked for ASR. Across FLEURS, error rates run from 1.41% to 63.76% — and 63.76 is Norwegian. Italian reaches 1.41, Spanish 2.20, German 2.73, French 3.16, English 3.35, Arabic 6.35 and Hindi 7.47; Japanese lands at 4.42 CER, Korean at 2.78 CER and Chinese at 8.62 CER. On the multilingual leaderboard's per-language means, Hindi is 9.15 OIWER against Dutch at 5.96 and German at 2.81.

Code-switching, which the card claims for all 40 languages, is similarly uneven: ASCEND Mandarin-English reaches 18.74% mixed error, MUCS Hindi-English 19.28% and MUCS Bengali-English 35.46%, with Monsoon Hindi at 9.40% OIWER. Translation quality spans a wide range too — FLEURS-to-English BLEU from 6.40 to 44.33, English-to-FLEURS from 8.70 to 42.65, and CoVoST2-to-English from 6.95 to 54.46. Robustness on the HALAS set is 31.50% WER, which is the number that should temper any plan built on accented or noisy field recordings.

The practical reading is that "40 languages" is a coverage claim you can plan capacity around and not a quality claim you can plan product around. Test the specific language and the specific recording conditions you care about on a few hundred of your own clips before committing. The model card's own fine-tuning case study is the template: on an Indian medical speech benchmark, out-of-the-box error was 30.8%, a prompt naming English and Latin script brought it to 23.8%, and adaptation on roughly 245 hours of data took it to 11.8% — about 62% fewer errors, on 14 GPU-hours of one GB300. Held-out speakers moved from 22.44% to 12.17%, a 45.8% relative reduction, while LibriSpeech test-clean slipped from 1.54% to 1.69%, the usual sign of a narrow adaptation.

What it costs to run, and where routing fits

NVIDIA has not published a price for the model. There is no per-minute rate on the research page or the model card, the weights are free under OpenMDW-1.1, and hosted access at launch runs through NVIDIA's own NIM endpoint plus several third-party platforms that set their own rates. So the cost question is a GPU question, and the numbers for it are unusually complete.

On one RTX PRO 6000 Blackwell Server Edition, the NVFP4 build with the DFlash2 draft processes up to 2,238 seconds of audio per second — a 30-second recording returns in 0.13 seconds — which puts 100 hours of audio under three minutes. That is the optimized NIM path, and NVIDIA states the open vLLM integration is roughly twice as slow. On an H100 at BF16 the card's RTFx figures are 410 with two-token speculation, 270 with eight, 340 with DFlash2 at K=7 and 260 at K=15. At the top of that range, one H100-hour covers about 410 hours of audio, so a thousand hours of transcription is roughly two and a half GPU-hours of compute. Whether that beats a hosted per-minute rate depends entirely on how saturated you keep the card: at a few hundred hours a month, an idle 80GB GPU is the more expensive option, and the arithmetic only flips once the queue stays full.

Adaptation is cheap by the standards of this class. NVIDIA reports 0.5 to 2.3 GPU-hours per task on a single GPU for new languages, domains or output styles, evaluated across 17 tasks, with LoRA on the language model and training on the encoder and projection. That is a small enough number that domain adaptation should be treated as part of deployment rather than a research project — the medical case above is 14 GPU-hours end to end.

One architectural note that matters for what comes next: this model emits text, and the summarisation, entity extraction and compliance tagging that follow a transcript are ordinary LLM work. That downstream stage is the part OrcaRouter covers — one API for 200+ models at 0% markup over provider list price, automatic failover when a provider degrades mid-pipeline, and a routing DSL for sending long transcripts and short snippets to different models without a second integration each time. We do not host Nemotron 3.5 Transcribe; it is a speech model you run on your own hardware or through NVIDIA's endpoint. The routing layer starts where the transcript ends.

Two models share the name, and only one is this one

A chunk of what has been written about this release in the last day is about a different checkpoint. NVIDIA also publishes nemotron-3.5-asr-streaming-0.6b, a small streaming ASR checkpoint built for on-device and CPU-oriented runtimes, where its quantised builds are small enough to ship inside an application. It is not this model, it does not do speaker attribution, and its speed and accuracy figures do not transfer. The 30B-A4B speech LLM is the one announced on October 28; the 0.6B streaming model is the one that runs on a laptop. If a comparison you are reading quotes sub-second latency or a one-gigabyte download alongside the 40-language claim, it has merged the two.

Who should act now, and who should wait

Act now if you run recorded meetings, calls or interviews in English or one of the twelve other diarisation languages, in files under an hour, and you already have 80GB-class GPU capacity or a hosted endpoint you are willing to pay for. The case is strongest where you are currently maintaining a two-model cascade: one checkpoint, one pass, speaker labels and word timestamps included, under a licence that permits commercial use. Evaluate it against your own audio before you retire pyannote — the far-field and accented-audio numbers above are the ones most likely to disappoint.

Wait if you need live streaming, because there is no streaming path; if your languages sit in the tail of that FLEURS range; if the recordings run past an hour; or if you have no GPU and no appetite for a hosted per-minute bill. Wait too if your stack is already on a newer transcriber that reads lower error on Artificial Analysis and does not need diarisation — the one-pass speaker attribution is the reason to move, and it is the only one.

The open questions worth watching are narrow and checkable. Independent full-precision evaluation is still thin; early reports from people running community 4-bit quantisations describe the usual slowdown and accuracy loss, which says more about the quant than the model. SGLang-Omni needs to reach feature parity before the non-vLLM path is a real alternative. And the thirteen diarisation languages are the number most likely to move first, because that is where the architectural claim is concentrated.