Nemotron 3.5 Transcribe: NVIDIA releases an open model that transcribes 40 languages, recognizes who is speaking, and translates, along with its training data

NVIDIA officially presented this Wednesday, October 28 at 4:14 PM Paris time, Nemotron 3.5 Transcribe, an open-weight speech recognition model. Its promise: replace the usual chain of several tools (transcription, then speaker identification, then translation) with a single model that you steer with natural-language instructions.

It transcribes 40 languages, including French, German, Spanish, Italian, Portuguese (including Brazilian), and Japanese, translates 31 of them to and from English, attributes speech to up to eight speakers in 13 languages, and accepts recordings up to one hour long.

A language model that listens

Technically, Nemotron 3.5 Transcribe grafts an audio encoder (about 800 million parameters, with a dedicated branch for voice recognition) onto Nemotron 3.5 Lightning, NVIDIA’s hybrid Mamba-Transformer mixture-of-experts language model. The whole thing weighs 34 billion parameters according to the model card, but activates only about 4.3 billion per token, making it usable on a single GPU. NVIDIA rounds to “30B” and “about 4B” in its posts.

Because it is a language model, you talk to it like an assistant: “transcribe with each speaker labeled,” “translate to English,” “what decisions were made in this meeting?” You can also feed it vocabulary, for example drug names, to avoid mistakes on rare terms. It gives word-level timestamps and can detect and mask personal data, a task present in small amounts in its training.

Open down to the data

This is the most striking point. The model is released under OpenMDW-1.1, a permissive license that allows any use, including commercial, provided the license and notices are retained upon redistribution. NVIDIA also makes available Granary v2, its multilingual speech dataset, the data preparation pipeline, and fine-tuning recipes with NeMo Speech. In a medical case study, fine-tuning reduced the error rate by 45.8% compared to the best prompt, according to NVIDIA.

For those who do not want to host it, NVIDIA cites access points at Baseten, DeepInfra, and Simplismart, without pricing in its documents.

What to keep in mind

The highlighted results come from NVIDIA. The company announces top spot among open models on the Hugging Face Open ASR Leaderboard in English (5.33% average errors in its post, 4.71% in the model card table, measured on October 12 across nine public datasets), a tie for first among open models at Artificial Analysis (2.8%), and first place on two speaker attribution benchmarks. These rankings move fast.

Other limitations stated in black and white: it is an offline model, not a real-time dictation service; audio question answering was only evaluated in English; and the record throughputs announced rely on optimizations not yet released as open source, the delivered version being about twice as slow according to NVIDIA.