Nvidia Nemotron Transcribe Explained: The Free Open AI That Turns Speech Into Text in 40 Languages (2026 Guide)

On October 28, 2026, Nvidia released a free open-weight model that could quietly change how every meeting, lecture, and phone call gets turned into text. Nvidia Nemotron Transcribe transcribes recorded audio in 40 languages, translates speech, labels up to eight speakers, and even answers questions about a recording — all from a single model you can run on one GPU. In this guide, we explain what it is, how accurate it claims to be, where it falls short, and whether it deserves a place in your workflow.

Table of Contents

What Is Nvidia Nemotron Transcribe?

Nvidia Nemotron Transcribe is a free, open-weight speech model that Nvidia released on October 28, 2026. It turns recorded audio into text in up to 40 languages, and it does much more than basic transcription: it can translate speech to and from English, label up to eight different speakers, answer questions about a recording, and follow plain-language instructions about how the transcript should look.

Think of it as a single checkpoint that replaces several separate tools. Instead of running one model for transcription, another for speaker labels, and a third for translation, the model handles all of those jobs from one prompt. You feed it an audio file, tell it what you want in ordinary words, and it returns the result in the format you asked for.

The release matters because transcription is one of the most practical uses of AI in everyday work. Meetings, lectures, interviews, medical notes, customer calls, and podcasts all produce hours of audio that nobody has time to replay. A strong open model changes the economics of that work: developers can build transcription into their own apps without paying a per-minute API fee or sending sensitive recordings to a third-party server.

Importantly, this is not a chatbot or a voice assistant. Nvidia Nemotron Transcribe works on recorded audio, not live streams. It listens to a finished file and produces text. If you need real-time captions during a live call, this is not the tool for that job, and Nvidia says so plainly. But for everything that can wait until the recording ends, it is built to be fast, accurate, and cheap to run.

Nvidia Nemotron Transcribe recording setup with a studio microphone

Alongside the model, Nvidia also released Granary v2, a huge open speech dataset covering 48 languages and more than 463,000 hours of audio. That dataset is the training foundation behind the model, and Nvidia opened it so other developers can train and fine-tune their own speech tools. We will come back to Granary v2 later, because it is a big part of why this release matters for the wider open-source community.

Nvidia Nemotron Transcribe at a Glance

  • Released: October 28, 2026, by Nvidia
  • What it does: transcription, translation, speaker diarization, and audio question-answering
  • Languages: transcription in up to 40 languages; translation between 31 languages and English
  • Speakers: labels up to eight speakers in a single recording
  • Audio length: handles recordings up to one hour long
  • Size: 30 billion total parameters, with about 4 billion active per token
  • Cost: free, open-weight model — you only pay for the hardware you run it on
  • Speed: Nvidia reports 100 hours of audio processed in under three minutes on one Blackwell server GPU
  • Key limit: recorded audio only — no live streaming or real-time captions

What Nvidia Nemotron Transcribe Can Actually Do

Most transcription tools do one thing: convert speech to text. Nvidia Nemotron Transcribe is designed as a promptable speech system, which means the same model changes its behaviour based on the instructions you give it. That flexibility is what sets it apart from older, single-purpose transcription engines.

Verbatim or Clean Transcripts, Your Choice

With Nvidia Nemotron Transcribe, you can ask for a verbatim transcript that keeps every “um,” pause, and false start, or a clean written transcript that reads like polished text. The difference matters: a lawyer reviewing a deposition wants every word exactly as spoken, while a product manager reading meeting notes wants the filler removed. Prompting the model lets you pick the style without switching tools.

You can also supply context vocabulary in the prompt. If your recording is full of product names, drug names, or company jargon, you can tell the model about those terms in advance. That is a practical answer to one of the oldest complaints about transcription software: it mangles the specialised words that matter most in professional settings.

Translation Between 31 Languages and English

Nvidia Nemotron Transcribe translates speech in 31 languages to and from English. That means a Spanish-language interview can come back as an English transcript, and an English briefing can be rendered in another language, all from the same audio file. For teams working across borders, that removes a whole step from the workflow — no separate translation pass, no second tool in the chain. Note the distinction: this translates a finished recording, whereas our Block Dessa guide covers live voice translation during a call.

Who Said What: Speaker Labels for Up to Eight People

Speaker diarization — the technical name for figuring out who spoke when — is built in. Nvidia Nemotron Transcribe can label up to eight speakers across 13 languages for diarization, and it reports the lowest error rate of the 14 systems tested on a 21-language speaker-attribution leaderboard. In plain terms, meeting transcripts come back with names or speaker tags attached, so you can follow a discussion without guessing who said what.

Nvidia Nemotron Transcribe turning a business meeting discussion into speaker-labelled text

This is where many older tools stumble. They transcribe the words fine but merge speakers together, or they split one person into two speakers every time that person changes tone. Nvidia’s benchmark claims suggest this model handles overlapping speech and multi-speaker sessions better than most open alternatives, though — as with all vendor benchmarks — real-world testing on your own audio is the only way to be sure.

Ask Questions About a Recording

Perhaps the most interesting capability is audio question-answering. You can ask Nvidia Nemotron Transcribe a question about the recording — for example, “What did the client say about the delivery deadline?” — and it answers from the audio itself, rather than making you search a transcript. This turns a long recording into something you can interrogate, which is a genuine time-saver for researchers, journalists, and anyone who reviews hours of interviews.

One caveat from the model card: audio understanding and spoken instruction following were evaluated only in English, so question-answering is strongest there even though transcription covers 40 languages.

How Nvidia Nemotron Transcribe Works Under the Hood

Under the surface, Nvidia Nemotron Transcribe is a mixture-of-experts model. The full checkpoint is about 34 billion parameters, built around Nvidia’s Nemotron 3.5 Lightning 30B-A3B language backbone, but only around 4 billion parameters activate for any given token. That sparse design is the key to its efficiency: the model carries a large amount of knowledge, but each word costs only a fraction of the full model’s compute.

The speech side combines a 600-million-parameter Transformer audio encoder, a roughly 99-million-parameter speaker encoder borrowed from Nvidia’s Nemotron 3 Diarization model, and a timing component that produces word-level timestamps. The language side is a hybrid Mamba-2 and Transformer mixture-of-experts design. Weights ship in BF16, with FP8 and NVFP4 quantized versions, plus an optional draft model for speculative decoding.

The context window is 65,536 tokens, which is part of how the model handles recordings up to an hour long. Training drew on roughly 1.72 million unique indexed audio hours across about 372 million examples, plus 5.6 billion unique text tokens, including Granary v2 data, multi-speaker meeting recordings, audio-understanding data, and synthetic speech designed to be aware of personal information.

For developers, the practical takeaway is simple: this is a serious, modern speech model, not a demo. It serves today through vLLM, and it can also run locally with NeMo-Speech.cpp. Hosted endpoints are available through providers such as Baseten, DeepInfra, and Simplismart, and there is an interactive demo on Hugging Face Spaces.

Benchmarks: What the Numbers Say (and What They Don’t)

Nvidia’s reported results place the model at or near the top of the open-weight field. These are company-reported figures, not independent test results, so treat them as a strong starting signal rather than settled fact. Still, the breadth of the claims is notable:

  • English transcription: first place among open-weight models on the Hugging Face Open ASR Leaderboard, at an average word error rate of 5.33%.
  • Multilingual accuracy: Artificial Analysis reports that the model tied Mistral’s Voxtral Small for first place among open-weight models at 2.8% AA-WER — while activating about 4 billion parameters compared with Voxtral Small’s 24 billion.
  • Speaker attribution: the lowest diarization error rate (18.26%) across 21 languages on a 14-system leaderboard, just ahead of Nvidia’s own dedicated diarization model at 18.38%.
  • Full-session English diarization: the lowest concatenated word error rate (32.4% cpWER) of 14 systems on VoiceArena’s benchmark, with the fewest dropped words — deletions of 8.32% of reference words versus a median of about 31%.
  • Independent evaluations: Ocular AI’s Converse-STT evaluation ranks it the top open model, and it also appears on Josh Talks’ Voice of India leaderboard for Indian languages.

The efficiency story is just as important as the accuracy story. Activating 4 billion parameters to match a model activating 24 billion means lower hardware costs, faster processing, and the ability to run on a single GPU. For a business transcribing thousands of hours of audio, that difference shows up directly in the monthly bill.

That said, benchmarks are measured on curated test sets. Accents, background noise, overlapping speech, and domain jargon all make real audio harder than benchmark audio. The honest way to evaluate Nvidia Nemotron Transcribe is to run it on your own recordings — a meeting, a call, a lecture — and compare the output against the tool you use today.

Speed and Hardware: From Edge Devices to Data Centers

One of the headline claims for the model is that it runs on a single GPU, from the Jetson AGX Thor robotics controller and the DGX Spark desktop chip all the way up to Blackwell data-center hardware. That range matters: the same model can power an on-device transcription feature in a robot or laptop, or a high-throughput service processing a company’s entire call archive.

On one RTX PRO 6000 Blackwell Server Edition GPU using NVFP4 precision, Nvidia reports processing 100 hours of audio in under three minutes, measured on 30-second clips at high concurrency. Even allowing for the usual gap between a vendor benchmark and a production deployment, that is an extraordinary throughput figure. It means batch transcription — converting a whole archive of recordings — becomes a minutes-long job rather than an overnight one.

For individual users, the more relevant point is the other end of the range: a model this capable running on a single chip opens the door to private, local transcription. Your audio never has to leave your machine. For doctors, lawyers, and journalists handling sensitive recordings, that privacy property is often worth more than any benchmark score.

Granary v2: The 463,000-Hour Dataset Behind the Model

Released alongside the model, Granary v2 is Nvidia’s open multilingual speech dataset, now covering 48 languages — up from 25 in the first version — and 463,062 hours of audio. Each recording comes with three transcript styles: one that keeps every filler word, one that converts spelled-out numbers into digits, and one that drops filler words entirely.

Why should non-developers care about a dataset? Because open datasets are what let the next generation of tools exist. Every startup that wants to build a better meeting assistant, a medical transcription app, or a multilingual call-center tool needs training data. By opening Granary v2, Nvidia lowers the barrier for all of them — and that competition eventually produces better, cheaper tools for end users.

Nvidia also published a medical fine-tuning case study that shows how the dataset and the model work together. Out of the box, the model scored a 30.8% word error rate on an Indian medical speech benchmark. A better prompt cut that to 23.8%, and fine-tuning with Granary-style data brought it down to 11.8% — with the tuning taking about 14 GPU-hours on a single GB300. That progression is a useful lesson: prompting helps, but fine-tuning on domain data is what turns a general model into a specialist.

Nvidia Nemotron Transcribe helping with medical dictation and clinical notes

Who Should Use Nvidia Nemotron Transcribe?

Meeting-heavy teams. If your work produces hours of meetings every week, the model can turn those recordings into searchable, speaker-labelled notes. Paired with a summarisation step, it replaces the ritual of replaying a call to find one decision. Transcribe is the newest member of Nvidia’s open-weight family, which we covered in our Nvidia Nemotron 3.5 guide, and finished meeting apps increasingly wrap models like this one into ready-to-use products.

Healthcare and medical documentation. Medical transcription is one of the two workplace footholds Nvidia explicitly targets, and the fine-tuning case study above shows why: domain vocabulary is where generic models fail, and this model can be adapted to it. Hospitals and clinics handling dictation should note the privacy angle too — a model that runs on local hardware keeps patient audio inside the building.

Podcasters and content creators. A podcast episode, a YouTube video, or an interview series all become searchable text with Nvidia Nemotron Transcribe. Clean transcripts improve accessibility, feed show notes, and give search engines text to index. Translation support means one recording can serve audiences in several languages.

Developers building voice products. If you are building a call-center analytics tool, a compliance recorder, a lecture-capture app, or a voice agent, this model gives you a free, high-quality speech layer to build on. Customisation through NeMo Speech lets you teach it your industry’s jargon, and the promptable design means one deployment can serve several output formats.

Researchers and journalists. Long interviews and field recordings become queryable. Instead of scrubbing through audio, you can ask what was said about a topic and get an answer grounded in the recording. For archival work, the batch speed — 100 hours in minutes on server hardware — changes what is feasible.

Nvidia Nemotron Transcribe transcribing customer support calls with a headset

Limitations You Should Know Before You Rely On It

No honest review skips the limits, and the model has several worth knowing:

  • Recorded audio only. It does not do live streaming or real-time captioning. If your use case is live, you need a streaming model instead.
  • Benchmarks are vendor-reported. The headline scores come from Nvidia’s own testing and partner evaluations. Independent testing on noisy, accented, real-world audio may tell a different story.
  • Audio Q&A is English-first. Audio understanding and spoken instruction following were evaluated only in English, so other languages may lag on the question-answering features even when transcription is strong.
  • Specialised vocabulary still needs work. The medical case study is encouraging, but it also shows the starting point: 30.8% errors before prompting and fine-tuning. Domain adaptation is a real project, not a switch you flip.
  • It is infrastructure, not an app. There is no polished consumer app here. Using Nvidia Nemotron Transcribe means running a model checkpoint, using a hosted endpoint, or building on top of it — it is a tool for developers and technical teams first.
  • Privacy still needs care. Running the model locally keeps audio private, but using a hosted endpoint sends your recordings to a provider. Check the provider’s data policy before uploading sensitive material.

Nvidia Nemotron Transcribe vs the Competition

The voice-AI field has been unusually busy in October 2026. Mistral shipped Voxtral 3 earlier in the month, ElevenLabs launched its Realtime models for instant speech in voice agents, and Salesforce released Launa-Omni. Nvidia Nemotron Transcribe enters that race with a different pitch: it is not trying to be a conversational voice agent. It is trying to be the best open transcription engine, and on that narrower goal its benchmark claims look strong.

Against Mistral’s Voxtral Small, the headline comparison is efficiency: a tie on Artificial Analysis’ multilingual word-error metric while activating roughly 4 billion parameters against 24 billion. Against Nvidia’s own Parakeet — the tiny 1.1-billion-parameter model praised for its size — Transcribe trades a larger footprint for far broader capabilities: translation, diarization, and question-answering in one checkpoint.

The closest comparison for most buyers is not another open model but the paid transcription APIs they already use. Those services charge per minute and keep your audio on their servers. Nvidia Nemotron Transcribe costs nothing in licence fees and can run on hardware you control. The trade-off is effort: you (or your vendor) must deploy and maintain it. For high-volume users, that trade is usually worth making; for occasional users, a finished app remains the simpler choice.

If you are following the voice-agent side of this story — models that talk back in real time rather than transcribe recordings — our comparison of Mistral Voxtral 3 and ElevenLabs Realtime covers that half of the October voice race in detail.

Nvidia Nemotron Transcribe running on a developer laptop for local speech processing

How to Try Nvidia Nemotron Transcribe Today

Getting started depends on how technical you are:

  1. Try the demo first. An interactive demo runs on Hugging Face Spaces, and hosted endpoints are offered by Baseten, DeepInfra, and Simplismart. Upload a recording and compare the output with your current tool.
  2. Run it locally. Developers can serve the checkpoint with vLLM or run it on-device with NeMo-Speech.cpp. A single modern GPU is enough to start experimenting.
  3. Prompt it deliberately. Test both verbatim and clean transcript styles, supply your domain vocabulary in the prompt, and check speaker labels on a multi-person recording before judging accuracy.
  4. Fine-tune for your domain. If you work in medicine, law, finance, or another jargon-heavy field, budget time for fine-tuning with data like Granary v2. Nvidia’s medical example shows this is where the biggest accuracy gains live.
  5. Keep sensitive audio local. If recordings contain patient, client, or unreleased business information, prefer a local deployment over a hosted endpoint.

For the official details, Nvidia’s announcement post on Hugging Face covers the architecture, benchmarks, and language coverage in full, and TechCrunch’s launch report gives a concise independent summary of what shipped.

Pricing: Free Weights, Real Hardware Costs

The model itself is free: the weights are open, with no per-minute fee and no subscription. That is the headline, and it is genuinely good news. But “free model” does not mean “free transcription.” You still need a GPU to run it, or you pay a hosting provider for endpoint access.

For an individual transcribing a few hours a week, a hosted endpoint or a local run on a capable PC keeps costs near zero. For a company processing thousands of hours, the cost becomes GPU time — which, given the reported throughput of 100 hours in under three minutes on server hardware, is far cheaper than paying a per-minute API for the same volume. The honest total cost depends on your volume, and for heavy users the open model wins by a wide margin.

The Verdict: A Serious Open Transcription Engine Arrives

Nvidia Nemotron Transcribe is one of the most useful open releases of October 2026 precisely because it is not chasing a demo. Transcription is unglamorous, high-volume, everyday work — and a free model that handles 40 languages, eight speakers, translation, and questions about audio, all on a single GPU, changes the cost and privacy calculation for everyone who does that work at scale.

It is not a finished consumer product, it does not do live captioning, and its benchmarks still need independent confirmation on messy real-world audio. But for developers, teams, clinics, creators, and researchers willing to run a model themselves, it is the strongest open transcription option available today — and the open Granary v2 dataset means the tools built on top of it will only get better.

If you already use a paid transcription service, the smart move is simple: run your hardest recent recording through the model and compare. The result will tell you more than any benchmark table can.

Frequently Asked Questions About Nvidia Nemotron Transcribe

Is Nvidia Nemotron Transcribe really free?

Yes. The model weights are open and free to use. You pay only for the hardware or hosted endpoint you run them on — there is no licence fee or per-minute charge from Nvidia.

How many languages does Nvidia Nemotron Transcribe support?

Transcription covers up to 40 languages, translation works between 31 languages and English, and speaker diarization covers up to eight speakers across 13 languages.

Can Nvidia Nemotron Transcribe do live captioning?

No. It is built for recorded audio, not live streams. For real-time captions you need a streaming speech model instead.

How accurate is Nvidia Nemotron Transcribe?

Nvidia reports first place among open-weight models on the Hugging Face Open ASR Leaderboard for English (5.33% word error rate) and a tie for first on Artificial Analysis’ multilingual metric. These are company-reported figures, so test it on your own audio before relying on it.

Where can I try Nvidia Nemotron Transcribe?

An interactive demo is available on Hugging Face Spaces, hosted endpoints are offered by Baseten, DeepInfra, and Simplismart, and developers can serve the checkpoint with vLLM or run it locally with NeMo-Speech.cpp.

Leave a Comment

Your email address will not be published. Required fields are marked *