NVIDIA Nemotron
lastmod 2026-10-30created 2026-10-301 edits
Nemotron is a family of artificial intelligence models developed by NVIDIA. It includes large language models and multimodal models for reasoning, computer programming, information retrieval, and agentic AI applications.
NVIDIA Nemotron 3.5 Transcribe
- NVIDIA Nemotron 3.5 Transcribe - supports 40 languages including Korean | GeekNews
- NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt
- Speech recognition
Nemotron 3.5 Transcribe provides transcription, speaker diarization, translation, and audio question answering in a single open-weight model, and its output format can be changed with natural-language instructions
- Supports transcription in 40 languages including Korean; for Korean, bidirectional translation with English and speaker diarization are also available. Handles recordings up to 1 hour and up to 8 speakers, and is not meant for real-time streaming
- Ranked first among open-weight models on the English Open ASR evaluation with an average WER of 5.33%, and also achieved the lowest error rate in the David AI and VoiceArena speaker diarization evaluations
- Granary v2 provides public data manifests covering 48 languages and 463,062 hours, plus a reusable cleaning pipeline, supporting training tailored to transcription formats and domain vocabulary
- In a medical speech case, fine-tuning lowered the WER on held-out evaluation speakers from 22.44% to 12.17%, a 45.8% relative improvement over the best prompt, though general English recognition performance dropped slightly