This week: NVIDIA released Nemotron 3.5 Transcribe, a free, open-weight speech model that transcribes 40 languages, labels who said what, translates, and answers questions about an hour-long recording, plus the open Granary v2 dataset behind it.

Elsewhere: How one model handles transcripts, speaker labels, and translation; how to try it without a data center; and what the leaderboard bragging omits.

The New Prime Signal

Anyone who has cleaned up an auto-generated meeting transcript knows the pain: half the words are right, nobody knows who said them, and the one name that matters is spelled three different ways. On October 28, NVIDIA released Nemotron 3.5 Transcribe, an open-weight model that "turns recorded speech into useful text," and you can tell it what kind of text you want in plain English. NVIDIA even demoed it on a scene from the 1940 comedy His Girl Friday, a film famous for people talking over each other at full speed. Bold choice. Most of us can barely follow that movie.

What happened: Transcribe combines four jobs in one checkpoint: transcription, speaker attribution, translation, and audio question answering. NVIDIA says it "transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13," handles recordings up to an hour long, and labels up to eight speakers. The weights are on Hugging Face under NVIDIA's OpenMDW-1.1 license, in full-precision and two compressed versions. Alongside it, NVIDIA published Granary v2, the open data pipeline that prepared part of the training data. Its public release covers 463,062 hours of speech across 48 languages, up from 25 languages in the first Granary.

Why it matters: Good transcription has mostly lived behind paid APIs, or in open models that only do one piece of the job. Here you can ask the same file for different things, like "Transcribe this recording and label each speaker" or "What decisions were made in this meeting?" It is also small where it counts: only about 4 billion of its 30 billion parameters are active per token, which is why NVIDIA says it can run locally on hardware like a DGX Spark or Jetson AGX Thor. On the Hugging Face Open ASR Leaderboard, NVIDIA says it ranks first among open-weight models for English, at 5.33% average word error rate.

The fine print: The leaderboard wins are real rankings, but NVIDIA picked which ones to highlight, and several are "among open-weight models," not against everything. It is built for recorded audio, "not live streaming," so it won't caption your call as it happens. Language support varies by task, and audio question answering is trained in English only. The big speed claim, 100 hours of audio in under three minutes on one GPU, used a quantized version NVIDIA calls "forthcoming." And Granary v2 ships manifests, not audio: you download the source recordings separately under each corpus's own license.

This is a genuinely useful release. One open model that knows who is talking, in a lot of languages, with the data recipe published next to it, is the sort of thing small teams can actually build on. It won't replace a human checking the quote before it goes in print, and NVIDIA says so itself. But "free and good enough to start with" covers a lot of podcasts, classrooms, and clinics.

The trick is that one model listens, keeps track of speakers, and takes instructions at the same time.

How it works: A speech encoder turns the audio into features, and speaker cues from NVIDIA's Sortformer diarization system ride along with them into a language model, so it can tag each turn while it writes instead of guessing afterward. You steer it with prompts. For a medical recording, NVIDIA suggests listing likely terms such as "HbA1c, metformin" and telling the model to use them "only when supported by the audio." When prompts aren't enough, you fine-tune. In NVIDIA's case study on recordings from Indian doctors and medical students, fine-tuning cut the word error rate on held-out speakers from 22.44% to 12.17%, a 45.8% relative drop, in about 14 GPU-hours. Granary v2 supplies the training targets for that kind of work, with three versions of each transcript: fully spelled-out verbatim, numbers as digits with the "uh"s kept, and a clean version without them.

Try it: The easiest route is the interactive demo linked from NVIDIA's announcement: upload a short clip and ask for speaker labels or a translation. Developers can grab the checkpoint from the model card and serve it with vLLM or SGLang-Omni, or use a hosted endpoint from Baseten, DeepInfra, or Simplismart. Start with something you already know well, like a podcast you recorded, so you can catch its mistakes. Then check every name, exactly as NVIDIA tells you to.