Nvidia launches open Nemotron 3.5 speech model for transcription, diarization and translation

Nvidia has launched Nemotron 3.5 Transcribe, an open-weight artificial intelligence model that combines multilingual transcription, translation, speaker diarization, and audio question answering into a single checkpoint. Available on Hugging Face and ModelScope alongside the Granary v2 data pipeline and fine-tuning tools, the model runs on a single Nvidia GPU ranging from DGX Spark to Blackwell data-center systems. [1][2]

The promptable system transcribes 40 languages, translates 31 languages to or from English, and identifies up to eight speakers across 13 languages. It handles recordings lasting up to an hour and supports word-level timestamps, though audio question answering is trained only in English. In a medical case study cited by Nvidia, fine-tuning cut the word error rate by 45.8% on held-out speakers compared to prompting alone. [2]

Rather than relying on disjointed multi-step pipelines, the model embeds speaker cues directly into the language model alongside acoustic features to label turns during generation. Nvidia reports that Transcribe achieved a 5.33% average word error rate on the Hugging Face Open ASR English leaderboard, placing it first among open-weight models, while posting an 18.26% error rate on David AI's 21-language diarization benchmark. [1][2]