NVIDIA Releases Open-Weight Nemotron 3.5 Transcribe Speech Model

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, speaker attribution, translation, and audio question answering in a single checkpoint. The model transcribes 40 languages, translates 31 languages to or from English, and attributes speakers in 13 languages. It supports recordings up to one hour long, labels up to eight speakers, and lets developers steer output format through natural-language prompts.

On public benchmarks, Nemotron 3.5 Transcribe ranks first among open-weight models on the Hugging Face Open ASR leaderboard for English with a 5.33% average word error rate. It also posts the lowest error rate on David AI's 21-language diarization leaderboard and ranks first in VoiceArena's 14-system comparison. Built on a hybrid Mamba-Transformer mixture-of-experts architecture, the model activates roughly 4B of its 30B parameters per token and runs on a single NVIDIA GPU, from Jetson AGX Thor and DGX Spark to Blackwell data-center systems.

Alongside the model, NVIDIA open-sourced Granary v2, a multilingual data curation pipeline used in training. The public manifests cover 463,062 hours across 48 languages and 16 source corpora, with aligned transcript variants, task labels, and quality metadata. Developers can run the pipeline on their own recordings or reuse the published manifests.

NVIDIA also released NeMo Speech fine-tuning tools and a medical case study. Fine-tuning on Eka Care's Indian medical speech benchmark reduced held-out word error rate from 22.44% to 12.17%, a 45.8% relative improvement, using about 14 GPU-hours on a single NVIDIA GB300.