Nvidia Nemotron Transcribe: Free AI Transcription in 2026

Nvidia just gave away a transcription model good enough to worry companies that charge for one. On October 28, 2026, the company released Nemotron 3.5 Transcribe, a free open-weight speech model that transcribes audio in up to 40 languages, translates 31 of them to and from English, and tells up to eight speakers apart in a single recording. If you transcribe client calls, interviews, or podcast episodes for a living, this is the rare free release that could change your monthly bill — with a few honest catches that decide whether it actually will.

Here is the short version. Nemotron 3.5 Transcribe is one checkpoint that does transcription, speaker labeling, translation, and audio question-answering, steered by plain-English prompts, running on a single Nvidia GPU. It handles recordings up to an hour long, returns word-level timestamps, and ships with Granary v2, a 463,062-hour open speech dataset and pipeline for teaching it your own vocabulary. It does not do livestreams. It does need real hardware and some setup patience. And for most freelancers, the smart move is not to rip out your paid tool tomorrow — it is to run this alongside it for a month and compare the transcripts on your own audio.

What Nvidia actually released

Most speech tools chain several steps together: detect speech, write it down, figure out who spoke, then summarize. Nemotron 3.5 Transcribe folds those jobs into a single model you steer with natural-language instructions. Ask it to transcribe a call and label each speaker, translate the audio into English, or tell you what decisions a meeting reached — the same recording can produce a different output depending on the prompt.

The headline numbers, all from Nvidia’s own release materials and covered by TechCrunch on October 28:

• 40 languages transcribed, 31 languages translated to or from English
• Speaker attribution in 13 languages, with up to 8 speakers per recording
• Recordings up to an hour long, with word-level timestamps
• Promptable output: you can supply jargon, names, and formatting instructions so terms come out spelled correctly
• Audio question-answering: ask follow-up questions about a recording in English, in the same pass

Under the hood it is a mixture-of-experts model with 30 billion parameters — about 34 billion counting its audio encoder, according to independent coverage — but only around 4.3 billion parameters are active for each token. That is the trick that makes a big model runnable on one chip: most of the model stays asleep while a small, relevant slice does the work. Nvidia says it runs from a Jetson AGX Thor edge device and DGX Spark desktop chip up to Blackwell data-center GPUs, with deployment options including vLLM, NeMo-Speech.cpp, and hosted endpoints from Baseten, DeepInfra, and Simplismart. Quantized versions for local RTX machines were described as coming soon at launch, so check the current model card before assuming your exact card is supported today.

The model ships under an open-weight license, with checkpoints available in BF16, FP8, NVFP4, GGUF, and DFlash2 formats — meaning you can run it at whatever precision your hardware allows. Alongside it comes Granary v2: 463,062 hours of speech data across 48 languages and 16 corpora, plus a reusable NeMo Curator pipeline. That dataset matters more than it first appears, because it is also the adaptation path: if the model mangles your industry’s vocabulary, the pipeline and fine-tuning recipes show you how to teach it.

On benchmarks, Nvidia claims the top spot among open-weight models on Hugging Face’s Open ASR Leaderboard for English at 5.33 percent average word error rate, and a tie on Artificial Analysis at 2.8 percent AA-WER while using about 4 billion active parameters against 24 billion for the model it ties. For speaker attribution it posts the lowest error rate on David AI’s 21-language diarization leaderboard and ranks first in VoiceArena’s 14-system comparison. Benchmarks are vendor-selected, so treat them as a signal, not a verdict — but the signal is strong, and the model is free, which makes testing it yourself cheap in every sense except your time.

The cost math: local vs paid, at three volumes

This is the part most coverage skips, and it is the part that decides whether you switch. Paid transcription pricing comes in two incompatible shapes: per-minute pay-as-you-go, and monthly subscriptions with capped minutes. To compare honestly, convert everything to what a month of audio actually costs.

Using per-minute rates from a current October 2026 pricing comparison:

Read that table carefully, because it tells three different stories. At one hour a month, a paid subscription is already overkill and a cheap pay-as-you-go service costs less than a coffee — local hosting is not worth the setup. At ten hours, pay-as-you-go starts to bite ($40–$150), and a subscription like Rev Essentials at $29.99 for 5,000 included minutes is genuinely good value. At fifty hours, pay-as-you-go becomes absurd ($200–$750), subscriptions look like bargains, and local hosting starts to look rational — if you already own the GPU.

That “if” is doing heavy lifting. The table’s $0 for local hosting is the marginal cost of electricity and your time, not the total cost. If you have to buy a GPU to run this, you are not saving money this year; you are prepaying years of transcription bills in hardware. If you already own a capable Nvidia card, or you already rent GPU time for other work, the marginal cost really is close to zero — and at that point every hour you transcribe locally is an hour you stop paying rent on.

There is also a speed claim worth noting. Nvidia says one RTX PRO 6000 Blackwell Server Edition GPU with NVFP4 precision can process 100 hours of audio in under three minutes — but that was measured on 30-second clips at high concurrency, so your results on real hour-long files will vary. Even at a tenth of that speed, batch transcription stops being a bottleneck and starts being something you run while you make coffee.

Who this is actually for

Freelancers transcribing client work

If you transcribe interviews, client calls, or research recordings, the workflow win is not just price. It is control. Your audio never leaves your machine, which matters when a client recording contains unreleased product details, a source who spoke off the record, or anything covered by an NDA. Paid services mostly promise not to train on your data; local hosting removes the question entirely. The speaker labels matter here too: an interview transcript without “who said what” is half a transcript, and eight-speaker support covers panel discussions and group calls that trip up simpler tools.

Podcasters and video creators

A podcast episode is the perfect Nemotron workload: recorded, not live, usually under an hour per segment, often with two to four speakers, and produced in batches. Word-level timestamps mean you can jump from a transcript line back to the exact audio moment for editing, and the translation feature gives you a first draft of subtitles or show notes in another language without a second service. If you already edit in Descript or a similar tool, keep it for editing — Nemotron replaces the transcription bill, not the editor.

Small teams with meeting archives

A team that records every standup, sales call, and training session generates exactly the kind of archive that gets expensive on per-minute pricing and messy on subscription caps. Local transcription turns the archive into searchable text at marginal cost, and the audio Q&A feature means you can ask “what did we decide about pricing in the March call?” instead of re-listening. The catch: someone on the team has to own the setup and the storage. A shared drive full of transcripts nobody maintains is not an archive; it is clutter.

Who should NOT switch yet

If you need live captions, this is not your tool — it is built for recorded audio, not livestreams, and Nvidia is explicit about that. If you transcribe in one of the 27 languages it does not cover, or you need speaker attribution in a language outside its 13, check the language list before investing setup time. And if you transcribe an hour a month and already have a tool you like, the honest answer is to stay put: the setup cost exceeds a year of $4 monthly bills.

The honest limitations

Free and open does not mean effortless. Here is what I would want to know before committing:

Hardware is the gate. “Runs on a single chip” is true and also misleading, because the chip in question is an Nvidia GPU with enough memory for a 30-billion-parameter model, even at reduced precision. Integrated graphics and older cards need not apply. Quantized versions were still coming at launch, so verify support for your specific card on the model card today, not in October’s headlines.

Recorded audio only. No livestreams, no real-time meeting bots, no live captioning. If your workflow is “join the call and transcribe as we speak,” tools like Otter.ai remain the right shape and Nemotron is the wrong one.

Accuracy varies by audio. Benchmark word error rates are measured on clean, standardized test sets. Your world has crosstalk, accents, background noise, and people talking over each other — the exact conditions where every transcription tool degrades. Nvidia’s own medical case study is instructive: on Eka Care’s benchmark, a better prompt cut word error rate from 30.8 percent to 23.8 percent, and fine-tuning reached 11.8 percent. Read that as good news (prompting and fine-tuning genuinely help) and as a warning (out-of-the-box accuracy on hard domain audio is not magic).

You are the support team. There is no help desk, no uptime guarantee, no “contact us” button. When a paid service breaks, you file a ticket. When a local model breaks, you read the docs. For a solo freelancer that trade is usually fine; for a team that needs someone to blame at 9 a.m. on a deadline day, it is a real consideration.

Privacy is a feature, not a promise. Local processing keeps audio on your machine, but it does not make you compliant with anything by itself. If you handle medical, legal, or financial recordings, your obligations around storage, access, and retention still apply — they just become entirely your responsibility.

Setup difficulty, honestly rated

I would rate the setup as moderate for a technical freelancer and hard for a non-technical one. You need to download a multi-gigabyte checkpoint in the right precision format, run it through vLLM or NeMo-Speech.cpp, and point it at your audio files. That is an afternoon’s work the first time, including the inevitable wrong-format download and the driver check. It is not a weekend project, and it is not a one-click install either. If you have ever set up a local AI model before — Stable Diffusion, Whisper, anything in that family — this will feel familiar. If you have never touched a command line, budget a full day and a patient friend, or use one of the hosted endpoints (Baseten, DeepInfra, Simplismart) and pay a small per-use fee instead of going fully local. The hosted route still beats most per-minute pricing while skipping the hardware question entirely.

A practical first test, whichever route you choose: take three recordings you already have transcripts for — one clean two-person call, one noisy group discussion, one accented or jargon-heavy session — and run all three through Nemotron and your current tool. Score them on the same three things: words wrong, speakers mislabeled, and jargon spelled correctly. That one afternoon tells you more than any benchmark table, including this article’s. If you are weighing a free local model against the wider paid-tool market, our AI tools for small business owners guide covers the broader landscape.

My verdict

Nemotron 3.5 Transcribe is the most useful free AI release for working freelancers in a month full of model launches, because it attacks a recurring bill rather than a hypothetical capability. It will not replace every paid tool: live transcription, the lowest-friction experience, and human-verified accuracy still belong to the services that charge for them. But if you transcribe recorded audio in volume, already own capable hardware, or care about keeping client audio local, the math is hard to argue with — at ten hours a month you save $40–$150 against pay-as-you-go, and at fifty hours the savings are several hundred dollars a month, every month, for a one-time setup afternoon.

The smart move this week is not a migration. It is a test. Download it, run your three hardest recordings through it, and let your own audio decide. Free tools earn their place the same way paid ones do: by being better on your work, not by being free in a headline.

If you want the wider landscape before you commit, our best AI transcription tools guide compares the paid field, and the official model card and release notes are the place to verify language support and hardware requirements before you download anything.

How to prompt it well: the skill nobody benchmarks

The most underrated feature of Nemotron 3.5 Transcribe is that it takes instructions in plain English — and the difference between a lazy prompt and a good one is measurable. Nvidia’s own medical case study showed a better prompt alone cutting word error rate from 30.8 percent to 23.8 percent on a hard benchmark, before any fine-tuning. That is a bigger improvement than switching between several paid tools.

In practice, a good prompt does four things. First, it names the output format: “Transcribe verbatim, label speakers as Speaker 1 and Speaker 2, and put timestamps every 30 seconds.” Second, it supplies the vocabulary: “The speakers are discussing Kubernetes, Terraform, and the client Acme Robotics — spell those terms exactly this way.” Third, it states what to do with the messy parts: “Mark overlapping speech as [crosstalk] rather than guessing words.” Fourth, it asks for the extra output you actually want: “After the transcript, list the three decisions made and who owns each one.”

Compare that with the default most people use — upload, click transcribe, accept whatever comes back — and you can see why two users can have wildly different opinions of the same model. The model is promptable in the same way a junior assistant is trainable: the more context you give it, the less cleanup you do later. Build a small prompt template for each recurring recording type (client call, interview, podcast, meeting) and reuse it; the ten minutes you spend once saves ten minutes on every file after.

Granary v2: the dataset is the long game

Most people will ignore Granary v2, and that is a mistake if you transcribe in a specialized field. The dataset — 463,062 hours across 48 languages and 16 corpora — is not just training data Nvidia used; it ships with a reusable NeMo Curator pipeline, which means you can build your own adapted training data in the same format and fine-tune the model on your vocabulary.

Here is a concrete scenario. A medical transcription freelancer finds the model handles general conversation well but mangles drug names and procedure terms. Instead of correcting every transcript by hand forever, she can assemble a few hours of her own corrected transcripts, run them through the same curation pipeline, and fine-tune. Nvidia’s Eka Care case study suggests the payoff is real: fine-tuning cut word error rate from 22.44 percent to 12.17 percent on held-out clips, with keyword error rate falling from 9.2 percent to 6.5 percent. Those are Nvidia-reported numbers on one benchmark, not a guarantee — but they show the direction: adaptation beats complaining about the base model.

For most freelancers, fine-tuning is a month-two project, not a day-one task. Start with prompting, which costs nothing. Graduate to fine-tuning only when you have a recurring vocabulary problem that prompting cannot fix, and a body of corrected transcripts to train on. The pipeline existing at all is what separates this release from a black-box API: when the model is wrong about your world, you have a documented path to fix it rather than a support ticket to file.

For the original announcement and the full technical detail, see TechCrunch’s coverage of the October 28 release and Nvidia’s own model release post on Hugging Face, which lists the language counts, deployment options, and benchmark figures used in this article.

Frequently Asked Questions

Is Nvidia Nemotron 3.5 Transcribe really free?

Yes — it is a free, open-weight model available on Hugging Face and ModelScope. There is no license fee for the model itself. Your costs are hardware (or hosted endpoint fees if you do not run it locally), electricity, and your setup time.

How many languages does it transcribe?

It transcribes up to 40 languages, translates 31 of them to and from English, and attributes speakers in 13 languages. Check the model card for the exact language list before relying on it for a specific language.

How many speakers can it separate?

Up to eight speakers in a single recording, with speaker labels included in the transcript. It also returns word-level timestamps so you can jump from text back to audio.

Can it transcribe live meetings or livestreams?

No. Nemotron 3.5 Transcribe is built for recorded audio only, not live streaming or real-time captioning. For live meetings, use a live transcription service instead.

What hardware do I need to run it locally?

A single Nvidia GPU, from a Jetson AGX Thor or DGX Spark up to Blackwell data-center cards. The model has 30 billion parameters but only about 4.3 billion active per token. Quantized versions for consumer RTX cards were coming soon at launch — verify your card on the current model card.

Is it more accurate than paid transcription services?

On public benchmarks it leads open-weight models, with 5.33 percent English word error rate on Hugging Face’s Open ASR Leaderboard. On your own audio, accuracy depends on noise, crosstalk, accents, and jargon — run your three hardest recordings through it and your current tool side by side before switching.