NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt
NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech-to-text model that transcribes, translates, and attributes speakers across multiple languages with promptable control.
NVIDIA Nemotron 3.5 Transcribe is an open-weight model designed to convert recorded speech into structured text with speaker attribution, translation, and audio question answering capabilities. It supports transcription in 40 languages, translation between 31 languages and English, and speaker diarization in 13 languages. The model processes audio up to one hour long and can label up to eight speakers, offering developers flexibility in output formatting and domain adaptation.
The model ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English with a 5.33% average word error rate, and ties for first on Artificial Analysis’ accuracy–active-parameter frontier at 2.8% AA-WER with approximately 4 billion active parameters. It also achieves the lowest diarization error rate on David AI’s 21-language benchmark and leads VoiceArena’s 14-system diarization comparison.
Granary v2, available on Hugging Face and ModelScope, provides a multilingual dataset of 463,062 hours across 48 languages, supporting prompt-level control for fine-tuning speech models. Developers can adapt Nemotron 3.5 Transcribe to domain-specific vocabulary and accents using NeMo Speech fine-tuning tools, as demonstrated by a medical case study where fine-tuning reduced word error rate by 45.8% on held-out speakers.
Nemotron 3.5 Transcribe can be deployed locally on RTX Spark, DGX Spark, or Jetson AGX Thor devices, or scaled on Blackwell data-center GPUs. Quantized versions are forthcoming, and the model is optimized for recorded audio rather than live streaming. Word-level timestamps and task-specific presets are supported, with recommended guidelines available in the model card.