Published: 29 October 2026
Nvidia released Nemotron 3.5 Transcribe on 28 October. The free, open-weight model turns recordings of up to an hour into text in 40 languages and translates 31 of them to or from English. It can also mark who is speaking, and it runs on a single Nvidia GPU. The catch is in the model card: Nvidia supports speaker labels in only 13 of those languages, and Dutch, Polish and Czech are not among them.
You choose the output with a prompt, and you can give it the names it will hear
Until now, Nvidia needed two models for a transcript with speakers. Parakeet wrote down the words, and Nemotron 3 Diarization, released in September, told the voices apart. Nemotron 3.5 Transcribe does both in one pass. A short prompt decides what you get back: a word-for-word transcript, a clean version, one with speaker labels and timestamps, or an English translation. You can also ask questions about a recording, such as which decisions were made, but Nvidia trained that skill in English only.
For work, the most useful option is the context hint. You give the model a list of names, product terms or drug names, and Nvidia says it should use them only when they occur in the recording. Nvidia’s own example shows how much jargon matters. On recordings of Indian doctors, the model got 30.8% of the words wrong out of the box. A more precise prompt brought that down to 23.8%, and a fine-tune on mostly synthetic speech to 11.8%.
The model works on recorded audio only, so it is no tool for live captions in a video call. It can tell apart up to eight speakers per recording.
Dutch is in Nvidia’s top tier for transcription, but not for speaker labels
The Nemotron 3.5 Transcribe model card sorts the languages into three quality levels. Dutch sits in the top level, which Nvidia defines as state of the art or close to it out of the box. Dutch also supports translation to and from English. In the column for speaker labels, however, Dutch has a dash, which the card explains as “not specified”. Czech, Polish and Slovak have the same profile.
German, French, Spanish and Italian do get speaker labels. Minutes of a Dutch meeting that show who said what therefore need a second step. That is either a separate speaker model or your own test of whether the labels hold up in Dutch anyway. Even where they work, the labels only separate voices within one recording, Nvidia writes. They do not establish who a person is.
Nvidia’s scores lead the open models, but Whisper is missing from the comparison
On Hugging Face’s public Open ASR leaderboard for English, the model ranks first among open-weight models with an average word error rate of 5.33%. Word error rate is the share of words a model gets wrong. Across seven languages, Nvidia’s model card reports 4.62%, against 6.49% for Alibaba’s Qwen3-ASR-1.7B. For Dutch, the card gives an average of 5.96% over three public test sets.
Those test sets mostly contain clear speech, much of it read aloud. Real meetings are harder. On long English recordings, the error rate in Nvidia’s own table rises to 10.98%. The announcement also leaves out OpenAI’s Whisper, the open model that many transcription tools are built on. Without that comparison, a test on your own recordings is the only way to find out whether a switch pays off.
An RTX 5090 can run it, and the licence allows commercial use
Nvidia says one RTX PRO 6000 Blackwell server card handled 100 hours of audio in under three minutes, in a test with short clips processed in parallel. That figure says more about capacity for a large archive than about the wait for one meeting.
- Model ID: NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, with about 30 billion parameters of which about 4 billion are active per step.
- Download: free on Hugging Face and ModelScope, with a demo on Hugging Face Spaces. The model card lists no regional limits, so you can also run it in the Netherlands.
- Licence: OpenMDW-1.1, a permissive licence from the Linux Foundation. The model card says the model is ready for commercial use.
- Hardware: one Nvidia GPU, from a DGX Spark or an RTX 5090 in the compressed NVFP4 format to an H100 in a data centre. Other chip makers are not in the list.
- Hosted: through Baseten, DeepInfra and Simplismart. Nvidia’s announcement names no prices.
- Audio: WAV, FLAC, MP3 or OGG, one channel, up to an hour per file.
Test it on 100 of your own recordings before you replace a transcription service
For organisations that record interviews, client calls or council meetings, the main gain is control. The model runs on a server you own, so recordings with personal data do not have to leave the building. That makes the privacy assessment simpler than for most transcription apps. Translation in the same pass helps teams that report to an English-speaking head office or work with foreign partners.
If you already transcribe Dutch or multilingual recordings and have access to an Nvidia GPU, start a test now. Collect about 100 recordings with checked transcripts, the minimum Nvidia advises for testing adapted versions. Include jargon, accents and people talking over each other, and run the same set through the tool you use today, Whisper included.
If you need live captions, or Dutch minutes with reliable speaker labels straight away, this model is a component to test rather than a finished product. If you prefer a hosted endpoint to your own GPU, first check where that provider processes the audio. Nvidia’s announcement does not say.
For more information, visit the official announcement of Nemotron 3.5 Transcribe on the Nvidia blog on Hugging Face.
