AI BriefingNVIDIA Releases Nemotron 3.5 Transcribe and Granary v2 for Multilingual Speech Processing
·2026.10.28 23:58
Key point
Open-weight model supports 40 languages and speaker diarization, with a 30B parameter architecture using 4B active parameters per token.
Details
- NVIDIA released Nemotron 3.5 Transcribe and the Granary v2 data pipeline, available on Hugging Face and ModelScope.
- The model achieves an average WER of 5.33% on the Hugging Face Open ASR Leaderboard for English, ranking first among open-weight models.
- Architecture utilizes a Speech encoder and Hybrid Mamba-Transformer MoE LLM with 30B total parameters and approximately 4B active parameters per token.
- Granary v2 expands to 48 languages and 463,062 hours of data, built on the NVIDIA NeMo Curator pipeline.
Performance and Capabilities
- Supports transcription for 40 languages, speaker diarization for 13 languages, and translation between English and 31 languages.
- Handles recordings up to 1 hour long with up to 8 speakers.
- Diarization performance includes a 18.26% DER on David AI DAI-ASR-I18N and 32.4% cpWER on VoiceArena Diarization-Bench.
Adaptation and Deployment
- Fine-tuning via NeMo Speech tools reduced WER by 45.8% in a medical case study using held-out speakers.
- Processing speed reaches under 3 minutes for 100 hours of audio on a single NVIDIA RTX PRO 6000 Blackwell Server Edition GPU using NVFP4 precision.
- Deployment options include vLLM or NeMo-Speech.cpp on single GPUs such as Jetson AGX Thor or DGX Spark.
- Limited to recorded audio; does not support real-time streaming.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Nvidia OpenShell: Secure Runtime for Autonomous AI Agent Fleets