NVIDIA on Wednesday said it released Nemotron 3.5 Transcribe, an open-weight speech model, and Granary v2 training data on Hugging Face and ModelScope¹.

The company described Transcribe as a single promptable checkpoint for transcription, speaker attribution, translation and audio question answering¹. NVIDIA said it transcribes 40 languages, translates 31 of them to or from English, attributes speakers in 13 languages, handles recordings up to an hour and labels up to eight speakers¹.

NVIDIA said the 30-billion-parameter model uses a hybrid Mamba-Transformer mixture-of-experts language model with about 4 billion parameters active per token, and that it runs on a single NVIDIA GPU¹. The post said the model is built for recorded audio, not live streaming¹.

On the Hugging Face Open ASR Leaderboard English track, NVIDIA said Transcribe is first among open-weight models at 5.33% average word error rate¹. It said Artificial Analysis ties the model for first among open-weight systems at 2.8% AA-WER¹. For speaker attribution, NVIDIA said it posted the lowest diarization error rate on David AI's 21-language track at 18.26% and ranked first of 14 systems on VoiceArena's English full-session comparison¹.

Granary v2, NVIDIA said, is an open curation pipeline whose public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1¹. The company said those figures describe the dataset release, not Transcribe's full training mixture¹.

In a medical automatic-speech-recognition case study on Eka Care's Indian medical speech benchmark, NVIDIA said fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt, from 22.44% to 12.17%¹. Hosted endpoints listed in the post include Baseten, DeepInfra and Simplismart¹.

Sources

1. Hugging FaceNVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt10:58 AM

Read the original at Hugging Face ↗