NVIDIA makes transcription style selectable for each recording
Nemotron 3.5 Transcribe offers downloadable speech-model weights and per-request control over transcript conventions, including speaker labels and timestamps that do not establish a speaker’s real identity.
NVIDIA released Nemotron 3.5 Transcribe on October 28 for transcription and speech translation in 40 languages, with recordings up to an hour long.
A natural-language instruction can request verbatim wording, written-form text, contextual terminology or speaker-attributed transcription without retraining. The model can return word-level timestamps with speaker labels. [1][2]
The change is useful because a meeting record, a readable transcript and a subtitle file may need different conventions from the same recording. The model card supplies weights and usage guidance under OpenMDW 1.1, while distinguishing within-recording speaker labels from a person’s real-world identity. Those labels are a segmentation aid, not biometric identification. [2]
Sources & details 2 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- NVIDIA — Nemotron 3.5 Transcribe
October 28 release and instructed transcription modes, languages and duration.
- NVIDIA — model card and weights
Operational limits, OpenMDW 1.1 terms and speaker-label identity limitation.
Images & figures
A request can select verbatim, written-form or speaker-attributed transcription from the same recording. Speaker labels distinguish turns within that audio; they do not identify a real person.
Euangelion · Source · Usage terms