AI news · Nvidia · Speech AI
Nvidia’s free Nemotron 3.5 Transcribe handles 40 languages
The open-weight model writes down, translates and labels speakers in one pass, and Nvidia says a single GPU can run it.
Nvidia released Nemotron 3.5 Transcribe on 28 October 2026, a free speech model whose weights anyone can download and run (TechCrunch). Nvidia says it writes down speech in up to 40 languages and can tell up to eight speakers apart. It translates between English and 31 languages, according to Nvidia. It works on recordings of up to an hour and cannot follow a live stream, TechCrunch reports.
Transcribe also handles speaker labels and translation itself, so one model run replaces the several steps a typical transcription setup passes audio through, SiliconANGLE reports. A written instruction, a prompt, tells it what to produce, such as a translation or the decisions taken in a meeting. It has 30 billion parameters, the learned values that make up a model, and uses 4 billion of them for any one request. Nvidia says it fits on a single GPU, from a Blackwell data-centre card down to a small DGX Spark or Jetson box, with the compression needed depending on the hardware. Granary v2, the free speech dataset behind the model, came out the same day.
What renting such a card costs is on our cloud GPU price sheet, read 24 October 2026. An RTX PRO 6000, a Blackwell card with 96 GB of memory, costs $1.08 an hour from Vast.ai’s cheapest verified host and $1.69 on Runpod’s Community Cloud.
At 16-bit precision, 30 billion parameters take about 60 GB, which fits in the RTX PRO 6000’s 96 GB. So a team already renting that card for video work can caption and translate ad voice-overs on it between video jobs, at the hourly rate it pays and with no per-minute transcription fee.
Sources
- TechCrunch: Voice AI in vogue: Nvidia releases new open transcription model. Checked
- SiliconANGLE: Nvidia releases all-in-one adaptable Nemotron 3.5 audio transcription and translation model. Checked