NVIDIA released Nemotron 3.5 Transcribe as an open-weight model. It handles transcription, speaker attribution, translation, and audio Q&A in one checkpoint. The model supports 40 languages for transcription and 31 for translation. It can process recordings up to one hour long. It labels up to eight speakers. Developers can guide output with natural language prompts. Granary v2 provides multilingual data and fine-tuning recipes for adaptation. Fine-tuning on medical data cut word error rate by 45.8%. The model leads the Hugging Face Open ASR Leaderboard for English. It ties for first among open-weight models on Artificial Analysis’ accuracy frontier. It tops two speaker-attribution benchmarks outright. Quantized versions are coming for local and edge deployment.
What changed
Previous Nemotron models did not combine transcription, speaker attribution, translation, and audio Q&A in a single checkpoint.
What it unlocks
Running one model for transcription, speaker labeling, translation, and audio Q&A with natural language prompts.
- 40 languages transcribed
- 31 languages translated to/from English
- 13 languages for speaker attribution
- up to 1 hour audio length
What you need to act on it
- access to Hugging Face or ModelScope
- hardware capable of running 30B parameter model
Sources