News
Nvidia's Nemotron 3.5 Transcribe: An Open-Weight Model for Hour-Long, Eight-Speaker Recordings, and What an Editor Should Test
ยท RenderBob team
Nvidia released Nemotron 3.5 Transcribe on 28 October, a 30-billion-parameter open-weight speech-to-text model for recorded audio. TechCrunch gives no scores or licence name.
Nvidia released Nemotron 3.5 Transcribe on 28 October, a 30-billion-parameter open-weight speech-to-text model for recorded audio. The benchmark claims come from Nvidia, and TechCrunch's report gives no scores or licence name.
What was released
TechCrunch reports that Nvidia released the model on Wednesday 28 October, together with version 2 of its Granary open-source speech dataset. The model transcribes speech in up to 40 languages and translates between 31 languages and English. It handles recordings up to an hour long, can tell up to eight speakers apart and works on recorded audio, not live streams. The report says the 30-billion-parameter model can run on a single chip, from Jetson AGX Thor and DGX Spark up to Blackwell datacentre GPUs. Nvidia says it beats rivals on several benchmarks, including Hugging Face's Open ASR Leaderboard. TechCrunch describes it as free and open-weight but does not name the licence. It lists Nvidia's existing Parakeet, Mistral's Voxtral 3, ElevenLabs' Realtime models and Salesforce's Launa-Omni among other recent voice releases.
Not the same as the June release
A separate, smaller, streaming model called Nemotron 3.5 ASR, at about 0.6 billion parameters, was described by Daily in June as built for voice agents. The names are close, so check which one a tool or tutorial means.
Why an editor might care
Transcripts drive rough cuts. Our 16 October piece on Claude Code in DaVinci Resolve listed rough cuts from timestamped transcripts among the tasks an assistant could drive, with misplaced markers and the judgement of what works left to the editor. A model that separates speakers across an hour of recording could help with interviews and panels. An open-weight model that runs on one chip also means footage need not leave your own hardware, although a 30-billion-parameter model on a DGX Spark is not a typical laptop job.
What to test
- Read the licence. It is not named in the report, and commercial terms matter.
- Run your own material. Benchmarks use clean speech. Try noise, accents, crosstalk and specialised names.
- Check speaker separation on a real panel. Count how often the model assigns a line to the wrong speaker.
- Check timestamps. A transcript is only useful for cutting if the times are accurate.
- Compare against your current tool on the same ten minutes.
- Price the hardware. Single-chip does not mean cheap.
What to watch
The licence, published scores on the Open ASR Leaderboard, independent tests on noisy and multi-speaker audio, and support in editing software.
RenderBob 3 and this story
RenderBob 3 makes the same local-or-cloud choice for generated clips that a local transcription model makes for speech. Keep footage and references on the Mac when they should not leave it, and use the cloud for the shots that need it, clip by clip.
More from the blog
- Sanxingdui Reportedly Took About 1.2 Million Yuan by Monday, Eight Times the Forecast, and Still Averaged Two Viewers a Screening
Newer reports put the first days of China's first all-AI feature well above the forecast we cited on 24 October, and still very small against its reported cost.
- ByteDance Opens a Short-Drama Studio With Script, Storyboard and "Regional Remake" Tools: What Is Announced, and the Likeness Question It Raises
ByteDance launched a short-drama studio in Jimeng and Dreamina on 26 and 27 October. A remake feature generates new visuals for another market, and reports do not say what happens to performers' likeness.