Nvidia Nemotron 3.5 Transcribe Covers 40 Languages [2026]
Reported by shattered.io, Hugging Face, The Decoder and 1 more outlets4 sources ↓
Nvidia has released Nemotron 3.5 Transcribe, an open speech recognition model that handles transcription, speaker recognition, and translation and takes its instructions as text prompts.
Sources
- shattered.ioRead ↗Nvidia Nemotron 3.5 Transcribe Covers 40 Languages [2026]
- Hugging FaceRead ↗NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt
- The DecoderRead ↗Nvidia's open Nemotron 3.5 Transcribe combines transcription, translation, and speaker recognition
- OpenToolsRead ↗Nemotron 3.5 Transcribe handles recordings, with task and language limits
Details
Nvidia released Nemotron 3.5 Transcribe, an open-weight model for recorded audio, on Hugging Face and ModelScope. It transcribes 40 languages, translates 31 to or from English, and identifies speakers in 13; recordings can be up to an hour long and include as many as eight speakers. Text prompts can specify output such as speaker labels, translation, or formatting for supplied technical terms, and the model also supports audio question answering and word-level timestamps.
The model has 30 billion parameters, with about 4 billion active per token. Nvidia says it can process 100 hours of audio in under three minutes on an RTX PRO 6000 Blackwell Server Edition at NVFP4 precision; it does not support real-time processing. On the Hugging Face English Open ASR Leaderboard, it scored 5.33% average word error rate and ranked first among open-weight models. Artificial Analysis reports 2.80% AA-WER and about 4 billion active parameters; The Decoder gives a figure of 2.81% and says Voxtral Small scored slightly lower at 2.77%.
Nvidia also released Granary v2, a speech dataset of about 844,000 hours across 48 languages, along with tools for adapting the model to specialized data. In a medical fine-tuning case study, Nvidia reports that word error rate on held-out speakers fell by 45.8% relative to the best prompt, from 22.44% to 12.17%; the training data was mostly synthetic. Language support varies by task, and the audio question-answering feature was trained in English, according to Nvidia.
Written from coverage as of
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh