NVIDIA opens Nemotron 3.5 Transcribe, a speech model that transcribes, translates and identifies speakers

Medium Highlight of the week Oct. 28, 2026 · NVIDIA / Hugging Face

Published in edition no. 1 of Noticiário do Joãozinho · Saturday, October 31, 2026, afternoon edition

Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese

A single open-weights checkpoint transcribes 40 languages, translates 31 of them to or from English, identifies speakers in 13 (up to eight people) and accepts recordings of up to one hour. It leads open models on the English Open ASR Leaderboard. It came with the Granary v2 data pipeline: in a medical case, fine-tuning cut the word error rate by 45.8%. There are quantized versions to run locally on RTX Spark, DGX Spark or Jetson AGX Thor.

Why it matters It offers commercial-grade transcription and diarization that is open and able to run locally. Before adopting it, check the model card to see whether your language is among the 40.

Original source: NVIDIA / Hugging Face · Model page (NVIDIA Research) — https://huggingface.co/blog/nvidia/nemotron-transcribe-granary

Full edition PDF (in Portuguese) Read the full edition