NVIDIA Nemotron 3.5 Transcribe: Free Transcription That Names Every Speaker

Most meeting transcripts fail in the same quiet way. The words are roughly right, but nobody can tell who said them, the speaker changes vanish in overlapping talk, and a medical term or product name comes out mangled. NVIDIA's new model is aimed squarely at that failure. Nemotron 3.5 Transcribe, released on October 28, 2026, is a free, open-weight speech model that transcribes 40 languages, translates 31 of them to or from English, attributes speakers in 13 languages, and labels up to eight speakers in a single recording.

It is not a consumer app. There is no signup page, no monthly plan, and no hosted dashboard from NVIDIA itself. It is a checkpoint developers download and run on their own hardware, or reach through third-party hosted endpoints. That distinction shapes everything about its cost, its limits, and who should actually use it. The weights are free. The computer to run them is not, and the model only works on recorded audio, never on a live stream.

The headline benchmark numbers are strong, and they come with the usual caveat: most of them are NVIDIA's own reporting from its announcement, or from leaderboards NVIDIA cites. The speaker-attribution results are the more interesting part, because that is where rival open models have been weakest. On two separate diarization benchmarks, Transcribe posts the lowest error rates of every system tested. If you build meeting tools, medical dictation, or call analytics, that is the number worth reading twice.

What Nemotron 3.5 Transcribe actually is

Transcribe is one promptable checkpoint that does four jobs: transcription, speaker attribution (diarization), translation, and audio question answering. Give it a recording and a plain-language instruction, and it changes what it returns. "Transcribe this recording and label each speaker" gives you a diarized transcript. "Translate this audio into English" gives you a translation. "What decisions were made in this meeting?" gives you an answer drawn from the audio. Word-level timestamps are supported, so a transcript can be clicked back to the exact moment a sentence was spoken.

Underneath, the model pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model. NVIDIA describes it as a 30-billion-parameter model with only about 4 billion parameters active per token. That sparsity is the whole engineering point: a big model that behaves, at inference time, like a much smaller one. Independent coverage counts the full checkpoint at roughly 34 billion parameters once the speech and speaker encoders are included, so treat "30B" as the language backbone figure and "about 4B active" as the number that decides your hardware bill.

Language support is uneven by design, and NVIDIA's own task matrix says so. All 40 languages can be transcribed, but only 31 translate to or from English, only 13 get speaker attribution, and audio question answering was trained in English. If your use case is a language outside the diarization set, you get a transcript, not a who-said-what transcript. Check the matrix for your language before planning a product around it.

Recordings can run up to an hour long, and the model labels up to eight speakers. Transcribe's eight-speaker ceiling is a deliberate trade, not a missing feature: it covers a meeting room, a panel, or a family call, but not a conference hall. And the recorded-audio-only limit is absolute. NVIDIA states it plainly in its announcement: Transcribe is built for recorded audio, not live streaming. If you need real-time captions, this is the wrong tool, and no prompt will change that.

The benchmarks, with the caveats attached

On transcription accuracy, NVIDIA reports first place among open-weight models on the Hugging Face Open ASR Leaderboard for English, at 5.33% average word error rate (WER), with API models excluded from that ranking. Lower WER means fewer errors. The same announcement reports 4.62% mean WER across seven languages and 10.98% on the long-form track, both the lowest among the open-weight models in its comparison. Those are NVIDIA's reported figures from its own model-card comparison, not an independent audit, and the English ranking explicitly excludes the proprietary APIs that might beat it.

Artificial Analysis, an independent tester, gives a second data point. Its speech-to-text index puts Transcribe tied for first among open-weight models at 2.8% AA-WER, matching Mistral's Voxtral Small while activating about 4 billion parameters against Voxtral Small's roughly 24 billion. On AA-AgentTalk, the held-out voice-agent speech set that makes up half of the AA-WER index, Transcribe scores 1.8% WER against 2.3% for the next-best open-weight model. That efficiency gap matters more than the tie: comparable accuracy at roughly a sixth of the active parameters is what lets the model run on a single GPU.

Speaker attribution is where Transcribe pulls away from the pack. On David AI's 21-language diarization benchmark, it posts the lowest diarization error rate of the 14 systems tested, at 18.26%, just ahead of NVIDIA's own dedicated Nemotron 3 Diarization model at 18.38%. On VoiceArena's English Diarization-Bench, which covers full sessions with overlapping speech, it ranks first of 14 systems at 32.4% concatenated minimum-permutation WER, and it drops the fewest words: deletions are 8.32% of reference words, against a median of about 31%. Overlapping speech is the hardest case in diarization, and it is exactly where most transcripts quietly lose a speaker. These are still benchmark results on curated test sets, not a guarantee for your noisiest conference call, but the lead is consistent across two independent benchmarks, which is rarer than a single leaderboard win.

What it costs, honestly

The weights cost nothing. Transcribe is released under the OpenMDW-1.1 licence, and NVIDIA's announcement presents it as ready for commercial or non-commercial use. There is no subscription, no per-minute charge from NVIDIA, and no credit system. That is the entire pricing story for the model itself, and it is genuinely unusual for a model at this accuracy level.

The bill arrives one step downstream, in hardware. Transcribe runs on a single NVIDIA GPU, from a Jetson AGX Thor edge chip or a DGX Spark desktop system up to Blackwell data-centre hardware. NVIDIA's own DGX Spark desktop starts at $4,999 for the 64GB model, as we covered in our DGX Spark pricing breakdown, and a cloud GPU rented by the hour varies by provider and region. No published per-hour rate for running Transcribe exists, because NVIDIA does not sell the compute. Budget for the machine, not the model.

NVIDIA's throughput claim shows what that machine buys. On one RTX PRO 6000 Blackwell Server Edition GPU, using the forthcoming NVFP4 quantized version, the model processed 100 hours of audio in under three minutes, measured on 30-second clips at high concurrency. Read that sentence carefully: it describes the forthcoming quantized version, and independent coverage notes that those serving optimizations are not yet part of the open-source release. The available checkpoint ships in BF16, with FP8 and NVFP4 variants, and quantized versions are described as coming soon. Your throughput today will be lower than the headline figure until those versions land.

For teams that do not want to self-host at all, hosted endpoints exist from Baseten, DeepInfra, and Simplismart, with an interactive demo on Hugging Face Spaces and ModelScope Studios. NVIDIA's announcement names the hosts but publishes no per-minute price for any of them, so a hosted cost comparison cannot be made from the announcement alone. Check the host's own pricing page before budgeting, and do not assume the hosted price inherits the weights' $0 price tag.

The comparison that does have published numbers is Mistral's Voxtral 3 family, which we reviewed in our Voxtral 3 review. Voxtral Small Transcribe 3 costs $0.004 per minute, or $0.24 per audio hour, on Mistral's API, and the Mini batch model costs $0.0035 per minute. That is the price of convenience: a hosted API with no GPU to manage, against Transcribe's $0 weights plus whatever your GPU costs. At 100 hours a month, Voxtral Small would bill about $24; a single rented GPU hour can exceed that, or fall below it, depending on the provider. The honest answer is that self-hosting wins at volume and on privacy, and loses on simplicity.

Granary v2: the dataset behind the model

Released alongside the model is Granary v2, NVIDIA's open speech dataset and curation pipeline. Its public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1. NVIDIA is careful to note that these figures describe the dataset release, not Transcribe's full training mixture or its sampling weights, so do not quote 463,062 hours as the model's training data.

What makes Granary v2 useful is not just its size but its formatting discipline. The pipeline produces three transcript forms for each recording: verbatim text with filler words, inverse-normalized text with disfluencies, and inverse-normalized text without them. It also produces short audio captions, code-switching restoration of loanwords, speech-grounded question-answer pairs, contextual biasing terms, and translations with automatic quality scores. Developers can run the NeMo Curator pipeline on their own recordings or authorized public corpora, or build on the public manifests, which carry entity hints, instruction pairs, provenance, and a per-record quality flag. Source audio itself is downloaded separately under each corpus's original licence.

That pipeline is the reason the model can be adapted to a domain, and NVIDIA demonstrates it with a medical case study. On Eka Care's open benchmark of recordings from Indian doctors and medical students, naming English and the Latin alphabet in the prompt cut full-benchmark WER from 30.8% to 23.8%. Fine-tuning brought it to 11.8% on the full set. On the cleaner held-out test, 2,764 clips from speakers the tuning never saw, fine-tuning cut WER from 22.44% to 12.17%, a 45.8% relative reduction, with keyword WER falling from 9.2% to 6.5% and semantic WER from 10.0% to 7.6%. The adaptation took about 14 GPU-hours on one GB300, using roughly 153 hours of synthetic speech plus 92 hours of public accented and medical recordings. The trade-off is real but small: general English LibriSpeech test-clean WER rose slightly, from 1.54% to 1.69%, after medical tuning.

How to get started

Try the model first in the Hugging Face demo or on ModelScope Studios before downloading anything. If it fits, download the checkpoint from Hugging Face or ModelScope and serve it with vLLM, which NVIDIA says supports it from day one, or SGLang-Omni. Local deployment with NeMo-Speech.cpp is described as coming soon, alongside the quantized versions, so laptop-class local use is a near-future promise rather than today's setup.

For prompting, start with NVIDIA's own advice: name the language and script in the prompt, supply domain terms as context before training anything, and use the model card's recommended presets. A medical prompt in the announcement shows the pattern: name the terms that might appear, tell the model to use them only when the audio supports them, and specify the written form for numbers. For domain errors that survive prompting, the NeMo Speech fine-tuning workflow pairs representative speech with explicit written targets, and a fine-tuning agent skill lets a coding agent run that workflow end to end.

One practical warning comes from NVIDIA itself: spot-check names, speaker changes, and answers against the recording. No transcription model, including this one, is reliable enough to quote a person verbatim in a legal or medical record without verification. The diarization lead narrows the error, it does not remove it.

Pros and cons

Pros: free open weights under a commercial-friendly licence; 40 languages with translation for 31; the lowest diarization error rates among open models on two benchmarks; promptable output formats, speaker labels, and word-level timestamps in one checkpoint; single-GPU deployment; and a reusable open data pipeline for domain adaptation.

Cons: recorded audio only, with no live streaming; audio question answering trained in English only; speaker attribution covers just 13 of the 40 languages; quantized versions and local NeMo-Speech.cpp deployment are still coming; the headline throughput figure describes a forthcoming quantized build; hosted endpoint prices are unpublished; and every benchmark figure is either NVIDIA-reported or drawn from a leaderboard NVIDIA cites, so independent testing on your own audio is still required.

Comparison table

For the wider meeting-tool field, see our roundup of the 7 best AI meeting note-takers in 2026, which covers the hosted tools most teams will compare against a self-hosted model like this one.

FAQ

Is Nemotron 3.5 Transcribe really free?

The weights are free under the OpenMDW-1.1 licence, with no subscription or per-minute charge from NVIDIA. You still pay for the GPU you run it on, or for a third-party hosted endpoint, whose prices are not published in NVIDIA's announcement.

Can it transcribe live meetings?

No. Transcribe is built for recorded audio files up to an hour long. It does not support live audio streaming, so real-time captions need a different model, such as Mistral's Realtime variant.

How many speakers can it tell apart?

Up to eight speakers per recording, across 13 languages for speaker attribution. Transcription itself covers 40 languages, so a recording in a language outside the diarization set will transcribe without speaker labels.

Do I need a data-centre GPU to run it?

No, NVIDIA says it runs on a single GPU, from Jetson AGX Thor and DGX Spark up to Blackwell data-centre hardware. The fastest throughput figures describe a forthcoming NVFP4 quantized version, so expect lower throughput from today's BF16 checkpoint.

Is it better than Voxtral Small?

On Artificial Analysis's index, the two tie for first among open-weight models at 2.8% AA-WER, with Transcribe using far fewer active parameters. On speaker attribution, Transcribe leads on the David AI and VoiceArena benchmarks cited in its announcement. Voxtral remains the simpler choice if you want a hosted API and live transcription, which Transcribe does not offer.

Key takeaways

• NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026, a free open-weight model that transcribes 40 languages, translates 31, and labels up to eight speakers.
• Its diarization lead is the story: lowest error rates on David AI's 21-language benchmark and VoiceArena's overlapping-speech benchmark among the systems tested.
• The weights cost nothing under OpenMDW-1.1, but the GPU to run them, or a hosted endpoint with unpublished pricing, is where the money goes.
• It handles recorded audio only, with audio question answering trained in English and speaker attribution for just 13 languages.
• Granary v2's 463,062-hour manifest and the medical fine-tuning case study show the model is built to be adapted, not just downloaded.

Checked and written on October 29, 2026, from NVIDIA's October 28 announcement. Transcribe is the rare open model whose best numbers are about who spoke, not just what was said, and that is why it deserves a place in any team that turns recordings into records.

Sources: NVIDIA announcement on Hugging Face (October 28, 2026); TechCrunch launch report (October 28, 2026); Unite.AI launch coverage; SiliconReport launch coverage; Artificial Analysis speech-to-text leaderboard as cited in NVIDIA's announcement.