AI × AEC · CURATED INSIGHT
NVIDIA Releases Open-Weight Audio Model Nemotron 3.5 Transcribe
Frontier AI models · AI creation · AI productivity
NVIDIA launched open-weight audio model Nemotron 3.5 Transcribe and the Granary v2 dataset, integrating transcription, translation, diarization, and QA for 40 languages and up to 8 speakers.
NVIDIA released Nemotron 3.5 Transcribe, an open-weight audio AI model with a hybrid Mamba-Transformer MoE architecture featuring 34 billion total parameters and 4.3 billion active parameters. It unifies transcription, translation, speaker diarization, and QA into a single step for up to 1-hour audio across 40 languages, with weights publicly available.
Key takeaways
- The model unifies transcription, translation, diarization, and audio QA, supporting custom prompts and pre-supplied terminology context.
- NVIDIA claimed top open-weight ranks on the Open ASR Leaderboard and concurrently released the Granary v2 dataset covering 48 languages.
Source：INSIDE Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.