What happened

Nvidia has released Nemotron 3.5 Transcribe, an open speech recognition model that follows plain-text instructions — and it does three jobs that normally take three separate tools. Hand it a recording and it'll write down what was said, label who said it, and translate it, all depending on what you ask for in the prompt.

The numbers are the interesting part. It transcribes 40 languages, translates 31 of them to or from English, and identifies speakers in 13. It handles recordings up to an hour long with up to eight people talking. Change the instruction and the same meeting becomes a speaker-labelled transcript, an English translation, or a summary of the decisions that were made. You can also pass it technical terms — drug names, product codes, client surnames — as hints, and Nvidia says the model is only supposed to use them if they actually appear in the audio.

That last detail is a quiet win. Most real-world transcription errors aren't grammar problems. They're names, acronyms and jargon that the model has never heard before.

One model, three jobs

Picture the old way of doing this: one app for the transcript, a second for speaker labels, a third for translation, and a human tidying up afterwards. Each handoff costs time and each tool has its own quirks.

Nemotron 3.5 Transcribe collapses that into one step. "Transcription" just means turning sound into text. "Speaker recognition" — sometimes called diarization — means telling apart Person A from Person B, even when they interrupt each other. "Translation" is self-explanatory. Building all three into a single model matters because the model hears the audio once, rather than three tools each guessing at a compressed copy of it.

A huge model that only wakes up part of its brain

The model has 30 billion parameters. In plain terms, parameters are the adjustable knobs a model learns during training — more knobs usually means more capability and more computing power to run. Thirty billion sounds like a data-center problem.

But only about 4 billion are active per processing step. Think of a big consulting firm with 30 specialists on staff: for any single client question, only four of them get pulled into the room. The rest stay at their desks. That's why Nemotron 3.5 Transcribe runs on a single Nvidia GPU, from the compact DGX Spark up to data-center hardware. Nvidia says it chews through 100 hours of audio in under three minutes on an RTX PRO 6000 Blackwell Server Edition at NVFP4 precision — a compressed number format that trades a little accuracy for a lot of speed.

The scores: 5.33% word error rate, first among open models

Word error rate is the simplest quality measure in speech recognition: out of every 100 words, how many come out wrong. Lower is better.

On Hugging Face's English Open ASR Leaderboard, Nemotron 3.5 Transcribe ranks first among open-weight models with an average word error rate of 5.33%, edging out open models from Alibaba and IBM. In Artificial Analysis' comparison it scores 2.81%, close to the best open model — Mistral's much larger Voxtral Small at 2.77% — while using roughly a sixth of the active parameters.

Speaker recognition is where the gap widens. VoiceArena tested 14 systems on conversations where people sometimes talk over each other, and Transcribe took first place. It also dropped the fewest words: 8.32% missed, against a median of about 31%. On David AI's benchmark, which measures who-is-speaking accuracy across 21 languages, it posted the lowest error rate — just ahead of Nvidia's own dedicated speaker model.

Granary v2: 844,000 hours of training audio, out in the open

Alongside the model, Nvidia is releasing Granary v2, a dataset of 844,000 hours of audio, plus tools for adapting the model efficiently to specialized fields such as medicine. That's the difference between a demo and something a clinic or a law firm can actually deploy in 2026: you fine-tune on your own vocabulary instead of hoping the base model guesses right.

What it means for you

You don't need a GPU cluster to care about this. Here's where it lands in ordinary life.

At home

Family calls, voice notes from relatives, interviews with elderly parents about their childhood — the stuff you mean to transcribe someday and never do. Feed an hour-long recording in and ask for a clean text version, or ask for a translated one if the call was in another language. Eight speakers is plenty for a big family get-together.

At work

This is the obvious one. A one-hour team meeting becomes a speaker-labelled transcript without a per-minute bill, and you can rerun the same file with a different instruction to get a decision summary instead. If your company works across languages, the 31-language translation covers a lot of ground — no more exporting audio and uploading it to a separate service.

Running a business

Customer calls are a goldmine nobody mines. With an open model you can turn every sales call and support ticket into searchable text, tag who said what, and look for the phrases that come up before people buy or cancel. Because it's open-weight and runs on modest hardware, the audio never has to leave your own machines — a real advantage if your customers care about privacy. If you want to clean up, summarize or repurpose those transcripts afterwards, free tools like MyKreaTool will save you a subscription or two.

Studying

Recorded lectures are the classic case for a speaker-aware model: you get the lecturer's explanation and the student questions clearly separated, plus a translation if the course is in a language you're still learning. The technical-term hints help too — feed in the module's glossary and the model will spell 'photosynthesis' right instead of guessing.

Creative work

Podcasters and video creators spend hours on transcripts for captions, show notes and clip hunting. Speaker labels mean you can pull one host's lines out instantly. You can also point the model at a rambling voice-memo session and ask it to summarize the decisions you made — a decent outline for a script.

Best AI tools saved weekly in our channel — @aigobySubscribe →

A side income

Transcription and subtitling services still sell. A freelancer who can turn around a clean, speaker-labelled, translated transcript in minutes instead of hours has a pricing advantage, and there's no per-minute API fee eating the margin. The same goes for anyone building a niche product for a specific industry — medical dictation, legal interviews, local-language media.

How to try it right now

The model is open-weight, so the free route is to download it and run it yourself. You don't need to write code to start, but you do need somewhere to run it.

1. Grab the model for free. Go to Hugging Face, open the Models tab, and search for "Nemotron 3.5 Transcribe." The weights are published there at no cost. Bookmark the model card — that's where the official instructions live.

2. Check your hardware. It runs on a single Nvidia GPU, but it's still a 30-billion-parameter model, so a laptop without a dedicated graphics card will struggle. If you don't have one, rent a cloud GPU by the hour, or use a DGX Spark if you're lucky enough to have one on the desk.

3. Write your instruction like a sentence to a colleague. Something like: "Transcribe this recording and label each speaker." Or "Translate this to English and note who said what." Or "Summarize the decisions made in this meeting."

4. Add hints for your jargon. List the names, drug names, product codes or acronyms you expect. The model is designed to use them only when they actually occur in the audio, so you won't get phantom mentions.

5. For a specialized field, fine-tune with Granary v2. Nvidia ships the 844,000-hour dataset and adaptation tools — the right path if you're processing medical or legal audio at volume.

6. Keep it under an hour per file. That's the current limit, along with eight speakers. Split longer recordings before you upload.

Upsides and what changes

• One tool instead of three. Transcription, speaker attribution and translation in a single pass means fewer uploads, fewer errors from stacked tools, and one bill instead of several.

• Open weights. You can run it on your own hardware, keep sensitive audio in-house, and adapt it to your vocabulary.

• Genuinely thrifty compute. Four billion active parameters out of 30 billion is why it fits on one GPU — and why 100 hours of audio takes under three minutes on serious hardware.

• Best-in-class speaker detection. An 8.32% missed-word rate against a 31% median in VoiceArena is not a rounding error. It's the difference between a usable transcript and a mess.

• Prompt-controlled output. Same file, different instruction, different deliverable. That flexibility is new for open speech models.

Limitations

A few honest caveats. The model handles files up to an hour with up to eight speakers — outside that, you're splitting and stitching. It transcribes 40 languages but only translates 31 and identifies speakers in 13, so "40 languages" isn't the same as "full support everywhere." Running it still requires a GPU, which most people don't have sitting around; renting one adds cost and setup friction the marketing doesn't mention. Speaker labels also get shakier when people talk over each other or a room is noisy — first place in a benchmark is not the same as perfect, and 5.33% word error rate still means roughly one wrong word in twenty on an English leaderboard. And fine-tuning for medicine or law is a project, not a checkbox.

Conclusion

Nvidia just made a serious, open, do-everything speech model that fits on one graphics card — and the speaker attribution, where most transcription tools quietly fall apart, is its strongest card. For entrepreneurs, marketers and creators, the practical takeaway is simple: the cost and hassle of turning audio into usable text just dropped again.

One action for today: open Hugging Face, search for "Nemotron 3.5 Transcribe," and read the model card. Ten minutes of reading will tell you whether it fits the recordings you already have piling up.