Miraki Mind

Nemotron 3.5 Transcribe: transcripts that never leave the building

NVIDIA just dropped an open speech model that transcribes, translates and tells speakers apart, in German, French and Italian too. What that means for interviews and subtitles, and where Swiss German stands.

Miraki MindAI-assisted editorial · 31 October 2026 · 6 min read

On 28 October 2026, NVIDIA released Nemotron 3.5 Transcribe, a speech model with open weights. It transcribes recordings of up to an hour, translates, identifies the language and tells you who speaks when, with timestamps for every word. According to NVIDIA, it covers 40 languages.

This matters to us because we work in several languages. Rapido exists in German and Swedish, 21 ZUAN in English and Chinese, and our website runs in four languages, Miraki Mind included. Wherever people speak, in interviews, voice-overs or dialogue, there is often a transcript at the start, and at the end you need subtitles or a version in another language.

Interview footage is sensitive. Real people from a company are talking, sometimes about things that aren't public yet. So a model that transcribes and translates that kind of recording on your own hardware hits exactly the question of where a brand's material ends up while we work with it. We haven't tested Nemotron 3.5 Transcribe yet.

What is actually new

Speech recognition with open weights has been around for a while. What's new is how much of it sits in a single model, and that you steer it with an instruction instead of new training.

Style on request: verbatim, as readable text, with speaker attribution or with extra detail. You set the style per request, in one sentence.
Feeding it names: you can send along a list of expected terms, such as people's names or product names. On the ContextASR-Bench benchmark, NVIDIA says this cuts the error rate on those names from 12.21 to 2.20 percent, a bit more than five times fewer.
Who speaks when: a built-in diarisation model returns speaker labels and timestamps for every word. The labels tell voices apart within a recording; they don't say who the person is.
Languages: the model card puts German, French and Italian in the top tier, "Leading", with transcription, translation to and from English, multiple speakers and switching languages mid-conversation.

The fine print matters: translation runs between English and each language. German straight to French isn't listed in the model card. And Swiss German doesn't show up anywhere, neither as a language nor as a variant.

The interview becomes a transcript, computed on your own hardware

What it means for a production

With an interview, transcribing is the step before the edit: finding the good lines, building the story from them, and at the end you need subtitles. A transcript with speakers and word timestamps is the raw material for that. The model card doesn't mention a ready-made subtitle format like SRT, though. Getting from transcript to subtitle file stays a separate step.

For four languages, that means: according to the model card, a German interview can be transcribed and translated into English. French and Italian go through English or through another tool. And machine translation doesn't replace a proofread by someone who actually speaks the language.

It won't run without a proper machine. The model has 34 billion parameters, about 4.3 billion of them active. NVIDIA names an H100 with 80 GB as the reference. The hardware list also includes the RTX 5090, but only in the compressed NVFP4 variant. It runs on Linux, and on Windows through WSL2. A Mac isn't on the list. There's a GGUF version for NVIDIA's local runtime NeMo-Speech.cpp; its instructions mention Metal, Apple's graphics interface, but don't say how well this model runs on it.

One more thing: the model works offline on finished recordings. Live transcription during the shoot isn't its job.

What it means for a brand

Local means the recording stays on your own box. For interviews with staff, confidential product details or customer voices before a launch, that's a real difference from a cloud service where the file gets uploaded first.

The licence is called OpenMDW-1.1. According to the licence text, you can use the model commercially, modify it and pass it on, and there are no conditions on the outputs. Anyone who redistributes the model files has to include the licence and notices. In the model card, NVIDIA says the model is ready for commercial and non-commercial use. What that means for a specific job is for a lawyer to clarify if in doubt, not this post.

And the model can get things wrong. The model card lists hallucinated speech, the wrong language and misattributed speakers. It also warns that the model may misrepresent a recording and still deliver a plausible transcript. When a line in a film runs under a CEO's name, that's not a detail.

A transcript that sounds plausible isn't yet a transcript that's right.

Our take and the limits

Nemotron 3.5 Transcribe is a serious candidate when interviews and footage shouldn't leave the building: openly licensed, German, French and Italian in the top tier, speakers and timestamps included. The top scores on the Open ASR leaderboards (4.71 per cent word error rate in English, 4.62 per cent multilingual) are NVIDIA's own figures at launch.

A few things are still open. How well does the model understand Swiss German, which the model card doesn't mention? The model card itself notes that the training data isn't shown to cover dialects and accents evenly. NVIDIA does say, though, that the model can be adapted to new languages and tasks on a single GPU. And how does it hold up in Swiss German against OpenAI's Whisper or xAI's Grok Speech to Text? We've already tested Whisper, and it won us over. And how much manual work is left between transcript and finished subtitle? That's what we want to see on real footage before we say more.

What is actually new

What it means for a production

What it means for a brand

Our take and the limits

Sources