October has been a very loud month for voice AI — Mistral's Voxtral 3, ElevenLabs Realtime, Salesforce's Launa-Omni — and NVIDIA just added the most practical entry of the lot. Nemotron 3.5 Transcribe, released October 28, 2026, is a free, ...
October has been a very loud month for voice AI — Mistral's Voxtral 3, ElevenLabs Realtime, Salesforce's Launa-Omni — and NVIDIA just added the most practical entry of the lot. Nemotron 3.5 Transcribe, released October 28, 2026, is a free, open-weight model that folds the four jobs a transcription pipeline usually stitches together — transcription, translation, speaker labels, and questions about the audio — into a single checkpoint you can prompt. The catch is stated plainly and early, because it decides whether this model is for you at all: it works on recorded audio only. No livestreams, no real-time captions. If your audio is already a file, though, this is the most interesting open speech release of the year so far.
Disclosure: no affiliate links in this article. I have not run Nemotron 3.5 Transcribe — the capabilities below are NVIDIA's claims from its October 28, 2026 announcement, not independently verified results.
What actually shipped
Nemotron 3.5 Transcribe is a 30-billion-parameter mixture-of-experts model that activates roughly 4 billion parameters per token, pairing a speech encoder with a hybrid Mamba-Transformer language model. That sparse design is the reason NVIDIA can claim single-GPU deployment across a wide range: Jetson AGX Thor, DGX Spark / RTX Spark, all the way up to Blackwell datacentre hardware. NVIDIA's headline speed claim — one RTX PRO 6000 Blackwell Server Edition processing 100 hours of audio in under 3 minutes — was measured on 30-second clips at high concurrency, so read it as a throughput claim under ideal batching, not as "your hour-long meeting transcribes in two seconds."
What one checkpoint does, according to NVIDIA:
Transcription in 40 languages, on recordings up to one hour long, with word-level timestamps.
Translation in 31 languages to and from English.
Speaker attribution (diarization) in 13 languages, labelling up to 8 speakers — so a meeting transcript can say who said what, not just what was said.
Audio question answering — ask questions about a recording and get answers from it. NVIDIA notes this was trained in English only, and that language support differs by task, so check the model card for your language before assuming every feature works everywhere.
Promptable output. One checkpoint, guided by natural-language instructions, can produce verbatim transcripts or cleaned-up text, and you can hand it a jargon list — product terms, medical vocabulary, financial terminology — so domain language survives transcription instead of being mangled. That last feature is the one meeting and medical tools will actually feel.
Alongside the model, NVIDIA released Granary v2, an open speech dataset and curation pipeline whose public manifest covers 463,062 hours across 48 languages — the raw material for anyone who wants to fine-tune the model on their own domain rather than just use it as-is. Availability: Hugging Face, ModelScope, GitHub via NVIDIA NeMo Speech, NGC and Build.Nvidia.com; weights under the OpenMDW-1.1 licence, which allows commercial use. Hosted endpoints come from Baseten, DeepInfra and Simplismart, with an interactive demo on Hugging Face Spaces and ModelScope Studios. Quantized versions for local RTX Spark, DGX Spark and Jetson deployment are promised soon, not yet shipped.
Why this matters for meeting and medical transcription
Meeting and medical transcription are two of AI's oldest, most established workplace footholds — and also the two where the pipeline has historically been the most cobbled together. A typical production setup chains a transcriber, a separate diarization model, a translation step, and a cleanup pass, each with its own failure modes and its own place where speaker labels drift out of sync with words. Nemotron 3.5 Transcribe's pitch is that the speaker cues reach the language model alongside the audio itself, so the labels are produced while the transcript is written, not bolted on afterwards. If that holds up in practice, it removes an entire class of "who said that?" errors — and the promptable jargon list attacks the other chronic pain, domain terms that generic models consistently get wrong.
The open-weight part matters just as much. Meeting audio and medical dictation are exactly the kinds of recordings many organisations cannot, or will not, send to a third-party API. A model you can run on your own single GPU, under a licence that allows commercial use, changes the compliance conversation entirely. That is the real audience here: not consumers looking for a nicer notes app, but developers building products where the audio stays in-house.
The benchmark claims, honestly stated
NVIDIA's announcement leans hard on leaderboards. Here they are, with the attribution they deserve:
First among open-weight models on the Hugging Face Open ASR Leaderboard (English). Note the qualifier: NVIDIA's chart for this ranking leaves API models out, so "first" means first among open weights, not first overall.
Tied Mistral's Voxtral Small at 2.8% AA-WER on Artificial Analysis's index — again, first among open-weight models — while activating roughly 4 billion parameters per token against Voxtral Small's 24 billion. If true, that is an efficiency story as much as an accuracy story: similar error, far less compute per token.
Lowest error on the David AI 21-language diarization leaderboard, with NVIDIA's own dedicated Nemotron 3 Diarization model in second place — NVIDIA is, in effect, its own closest rival here.
First in VoiceArena's 14-system comparison of speech models.
All of these are vendor claims and leaderboard snapshots, not hands-on results. Leaderboards reward the test set; your meeting room, your accents, your overlapping speakers and your bad microphones are a different test entirely. The honest way to read them: Transcribe belongs in the same conversation as Voxtral and the best Whisper-class options, and it is cheap enough per token to be worth a head-to-head on your own data. It is not proof that it wins that head-to-head.
How it compares, without invented numbers
Versus Mistral Voxtral: Voxtral is the closest rival and the natural comparison — a strong open-weight speech model from earlier this month, and the one Transcribe ties on Artificial Analysis's index. Transcribe's claimed edge is efficiency (far fewer active parameters) and the all-in-one design: speaker labels and audio Q&A in the same checkpoint rather than a separate step. Voxtral's edge, if you already have it working, is incumbency: a model in production beats a model on a leaderboard. Run both on the same recordings before switching.
Versus Whisper-class options: Whisper remains the default open transcription workhorse, with the broadest tooling ecosystem and the most community fine-tunes. Transcribe's pitch against it is not raw transcription alone but the bundle — diarization, translation, prompting and Q&A in one place, plus a much larger model behind it. The trade-off is ecosystem maturity: Whisper has years of integrations; Transcribe has days. If your pipeline is Whisper-shaped and working, Transcribe is a candidate to evaluate, not an automatic upgrade.
Versus hosted APIs: If you do not care where audio goes, a hosted transcription API is still easier — no GPU, no deployment, no fine-tuning. Transcribe wins where audio must stay private, where per-minute API costs scale badly, or where you need to adapt the model to your own jargon. It loses on convenience, full stop.
The limit that matters: recorded audio only
This deserves its own section because it is the easiest way to waste a week. Nemotron 3.5 Transcribe does not do livestreams and does not do real-time captions. If you are building live meeting captions, a live broadcast tool, or anything where the transcript must appear while someone is still talking, this model cannot do the job — look at ElevenLabs Realtime or a streaming ASR setup instead. Transcribe is for the file you already have: the recorded meeting, the dictation, the archived call, the interview tape. Within that boundary it is generous — up to an hour per recording, word-level timestamps, eight speakers. Outside it, it is simply the wrong tool, and no benchmark changes that.
Who should use it — and who should skip it
Try it first if: you build meeting, medical, legal or support transcription on recorded audio; you need speaker labels and domain vocabulary to survive; your audio cannot leave your infrastructure; or you want to fine-tune on your own data with Granary v2 as a starting point. Start with the Hugging Face Spaces demo, then the hosted endpoints, then self-hosting.
Skip it if: you need live captions or livestream transcription; you want a consumer notes app rather than a model to build with; or your language or task is one where NVIDIA itself flags limited support — check the model card before committing.
Wait if: you need the quantized local versions for RTX Spark, DGX Spark or Jetson — those are coming soon, not shipped — or you want independent, hands-on verdicts from developers who have run it on real, messy audio. Those will arrive over the next few weeks.
What to try first
Try the demo on your worst recording, not your best. Overlapping speakers, accents, background noise, medical or financial jargon — the demo on Hugging Face Spaces or ModelScope Studios is free, and a clean studio clip tells you nothing.
Test the jargon list on your domain terms. Give it your product names, drug names or ticker terms and check whether they survive verbatim. This is the feature most likely to decide whether Transcribe beats your current setup.
Check speaker labels on a real multi-speaker meeting. Diarization is where Transcribe claims its strongest lead — verify it on your own meetings, where people interrupt each other, before trusting the leaderboard.
Compare cost per hour, not cost per token. If you self-host, measure throughput on your hardware with your clip lengths; NVIDIA's 100-hours-in-3-minutes figure was measured on short clips at high concurrency and will not match a single long file.
Do not rebuild your pipeline on day one. Run Transcribe in parallel with your current transcriber for a week, diff the outputs, and let your own error rate — not a leaderboard — make the decision.
Sources: NVIDIA on Hugging Face — Nemotron 3.5 Transcribe and Granary v2 (October 28, 2026) · TechCrunch — Voice AI in vogue: Nvidia releases new open transcription model (October 28, 2026) · Unite.AI — NVIDIA Launches Open-Weight Speech Model and 463,062-Hour Audio Dataset (October 28, 2026)