NVIDIA releases Nemotron 3.5 Transcribe across 40 languages
The open-weight transcription and speaker-attribution model arrives alongside the Granary v2 pipeline covering 463,062 hours of multilingual speech.
Image: Nvidia
NVIDIA introduced Nemotron 3.5 Transcribe, an open-weight model engineered for speech transcription, translation, and speaker attribution [1]. According to the release announcement, the system provides transcription coverage across 40 languages and speaker attribution across 13 languages [1]. The architecture handles audio recordings up to one hour in length and can label up to eight distinct speakers per session [1].
Alongside the model, NVIDIA detailed its Granary v2 data curation pipeline [1]. The public manifest associated with Granary v2 encompasses 463,062 hours of multilingual speech drawn from 48 languages [1]. This curated corpus supplies the training data foundation for the open-weight release, aiming to help developers build and adapt specialized voice pipelines without proprietary barriers [1].
The release adds to NVIDIA's growing portfolio of developer-facing models and multilingual machine learning resources [1]. By publishing open model weights along with the underlying curation details, the release allows teams to inspect the training pipeline directly and customize transcription tasks for multilingual enterprise workloads [1].
Questions and answers
What languages does Nemotron 3.5 Transcribe support?
The model supports transcription across 40 languages and provides speaker attribution for 13 languages.
What audio limits apply to the model?
Nemotron 3.5 Transcribe can process recordings up to one hour long and track up to eight speakers.
What is Granary v2?
Granary v2 is a speech data curation pipeline whose public manifest covers 463,062 hours of audio spanning 48 languages.
Companies & investors in this story
Sources
- 1.NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt — Hugging Face Blog, Oct 28, 2026