NVIDIA announces Nemotron 3.5 Transcribe for multi-speaker transcription

Summary

NVIDIA has released Nemotron 3.5 Transcribe, featuring a 34-billion parameter (4 billion active) MoE architecture. The model processes up to one hour of audio in 40 languages within a single checkpoint, offering speaker diarization, contextual prompting, and translation capabilities. It is released as open weights under the OpenMDW-1.1 license. Among open models, it ranks first on the Open ASR Leaderboard for English. On the Artificial Analysis WER index, it ties with Voxtral Small Transcribe.

Why it matters

By integrating transcription, speaker attribution, translation, and speech understanding into a single promptable checkpoint, the model eliminates the need for multi-model pipelines that are still common in production voice stacks. Its efficiency, achieving top-tier WER scores with only 4 billion active parameters, enables high-quality multilingual transcription with significantly less hardware compared to dense models with over 30 billion parameters. The simultaneous release of the Granary v2 training dataset also lowers barriers for fine-tuning and independent research.

Highlights

• 34B total / 4B active parameter MoE architecture processes 1 hour of audio in 40 languages with built-in speaker diarization
• Lowest WER among open models on the Open ASR Leaderboard (English) and Artificial Analysis index; ties with Voxtral Small Transcribe
• Open weights (OpenMDW-1.1); Granary v2 dataset released; quantized versions are planned but not yet available

Critical take

In NVIDIA's own tiering, Japanese is classified as 'Production-ready' rather than the top-tier 'Leader.' Additionally, the model does not outperform Voxtral Small Transcribe in WER comparisons but ties with it. Therefore, the claim of being the 'best open model' carries a narrower scope than the headline implies. Furthermore, OpenMDW-1.1 is a restricted open-weights license, not a permissive open-source license. The promised quantized versions are still only in the 'planned' stage. This limits immediate deployment flexibility for users with constrained resources.