Nemotron 3.5 Transcribe is NVIDIA's open-weight speech model that turns recordings into text and can also translate, tell speakers apart and answer questions about the audio.
In plain words
One model does four jobs: writing down speech, labelling who spoke (diarization), translating, and answering questions about a recording. NVIDIA says it transcribes 40 languages, can translate 31 of those into or out of English, and labels speakers in 13. It accepts recordings as long as one hour, with as many as eight speakers. You steer it with plain instructions. In a medical test, fine-tuning on new speech cut errors by 45.8% versus the best prompt. It was trained to answer questions about audio in English.
A quick example
Imagine recording a team meeting held in two languages. You ask the model to label each speaker, then to translate the whole thing into English, then to list what was decided.
Where you'll meet it
- Meeting and call transcripts with speaker labels.
- Translating recorded speech to or from English.
- Medical dictation with correctly spelled terms.
Related
Sources
- DocsNemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt, NVIDIA (Hugging Face blog) · read 28 Oct 2026
- DocsNVIDIA-Nemotron-3.5-Transcribe-30B-A4B, NVIDIA (Hugging Face) · read 28 Oct 2026
Checked by WhatIsAI against the sources above · 28 Oct 2026. Spotted an error? Report it.