NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt
Summary
- NVIDIA has released Nemotron 3.5 Transcribe, an open-weight, promptable multilingual speech model supporting 40 languages and audio Q&A.
- The model pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model, using roughly 4B active parameters out of 30B.
- It ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English and leads multiple diarization benchmarks.
- NVIDIA also released Granary v2, an open data curation pipeline and dataset covering over 463,000 hours of audio across 48 languages.
- The model and curation pipeline can be adapted using NeMo Speech fine-tuning tools to significantly lower word error rates on domain-specific vocabulary.
Why it matters
Nemotron 3.5 Transcribe sets a new performance bar for open-weight speech models by combining promptable multilingual transcription, translation, and speaker attribution into a single efficient checkpoint. Its release alongside the Granary v2 data pipeline gives developers robust, adaptable tools for building custom speech applications.