NVIDIA Opens Weights of Nemotron 3.5 Transcribe, a Promptable Speech Model Covering 40 Languages
One checkpoint handles transcription, speaker attribution, translation and audio Q&A across 40 languages, with best-in-class open-weight accuracy claims.
NVIDIA has released the weights, training dataset and technical report for Nemotron 3.5 Transcribe, a 30B-A4B speech-augmented language model announced on October 28. The model combines a frozen Transformer ASR encoder with diarization from Nemotron 3 and a Nemotron 3.5 Lightning hybrid Mamba–Transformer MoE backbone with roughly 4.3B active parameters. A single checkpoint handles verbatim, written-form, contextual and speaker-attributed transcription, speech translation between 31 languages and English, language identification, per-speaker word timestamps, and audio question answering — output style controlled through natural-language instructions without retraining.
NVIDIA reports a mean word error rate of 4.71% on the English Open ASR Leaderboard, the best open-weight result, and 4.62% multilingual, while contextual entity lists cut WER up to sixfold on ContextASR-Bench. The model supports recordings up to one hour with up to eight speakers, runs from Jetson AGX Thor-class devices to data centers, and is served by vLLM and SGLang from day one.
The accompanying Granary v2 dataset spans about 844,000 hours of audio across 48 languages, and NVIDIA reports a 45.8% relative WER reduction after fine-tuning on medical data, evidence of domain adaptability.