Overview
- Nvidia publicly released Nemotron 3.5 Transcribe, a 30‑billion-parameter mixture-of-experts model that transcribes 40 languages and translates 31 of them to or from English.
- The model accepts pre-listening prompts so users can set domain context, speaker labels, formatting preferences, and request cleaned or verbatim transcripts or audio Q&A.
- Nvidia says Nemotron ranks at the top of open-model ASR and speaker-attribution leaderboards and published a fine-tuning case study that cut medical-transcription word error rate from 22.44% to 12.17% after about 14 GPU-hours on one GB300.
- Nemotron is built for recorded audio up to one hour, can attribute up to eight speakers in supported languages, runs on a single GPU on hardware from Jetson AGX Thor to Blackwell, and does not support live streaming.
- Nvidia also released Granary v2, a NeMo Curator-based pipeline with public manifests covering roughly 844,000 hours of speech across 48 languages and 19 corpora to speed developer customization and domain adaptation for meeting and medical workflows.