Nvidia Releases Nemotron 3.5 Transcribe: One Open Model for Transcripts, Translation and Who Said What

Nvidia has released Nemotron 3.5 Transcribe, an open model that turns recorded speech into text, translates it, and labels which speaker said what, all in one. It transcribes 40 languages, translates 31 of them to or from English, and identifies speakers in 13. It handles recordings up to an hour with as many as eight speakers, and you can steer it with a text prompt, for example to get a summary or to hint at technical terms.

The model has 30 billion parameters in total (the learned settings inside it), but only about 4 billion are active at each step, which keeps it quick and light. It runs on a single Nvidia GPU, from the small DGX Spark desk machine to data centre cards. Nvidia says one server card processed 100 hours of audio in under three minutes. On the Hugging Face Open ASR English leaderboard it scores a 5.33% average word error rate, first among open-weight models. Word error rate is simply the share of words transcribed wrongly, so lower is better. On a test with overlapping speakers it missed 8.32% of words, against a median of about 31% across 14 systems. These are mostly Nvidia-reported or leaderboard figures.

Nvidia also released its training data, Granary v2, about 844,000 hours of speech in 48 languages, along with tools for building similar datasets. Fine-tuning (teaching the model your specific material) helped in one example: error on Indian doctors’ and students’ medical speech fell from 23.8% to 11.8%, in about 14 GPU hours.

What’s behind it: transcription used to need separate tools for text, translation and speaker labels, each adding errors. Bundling them is convenient, and publishing the data lets others build on the work. For Nvidia, open speech models also give people more reasons to buy its hardware.

Here is the fair caveat list. It works on recorded audio only, with no real-time mode. Feature support varies by language, and question answering about audio was trained only in English. Fine-tuning slightly hurt general English accuracy, and a specialised medical model handled brand-name drugs better. The article does not state the licence, so check the Hugging Face page before commercial use. Nvidia itself advises spot-checking names, speaker changes and answers.

What this means for you: if you just want to transcribe a meeting, you will not run this by hand today, but expect apps to adopt it. If you are more technical and have a capable GPU, try the free demo on Hugging Face with your own recordings first. For anything sensitive, such as legal or medical audio, keep a human in the loop.

Sources

• The Decoder: Nvidia’s open Nemotron 3.5 Transcribe