NVIDIA releases Nemotron 3.5 Transcribe open-weight model

NVIDIA has made Nemotron 3.5 Transcribe available as an open-weight model on Hugging Face and ModelScope. The system transcribes audio in up to 40 languages and translates 31 of those languages to or from English. It processes recordings up to an hour long and attributes speakers in 13 languages. The model can label up to eight speakers in a recording. It also supports word-level timestamps for transcription output and audio question answering, a feature trained specifically in English.

The model contains 30 billion parameters, with only 4 billion active per step. This architecture allows it to run on a single chip, such as the Jetson AGX Thor or DGX Spark. On an RTX PRO 6000 Blackwell Server Edition at NVFP4 precision, the model processes 100 hours of audio in under three minutes. NVIDIA also released Granary v2, a dataset that includes recordings in several transcription styles, including verbatim and cleaned-up versions.

NVIDIA tested targeted fine-tuning on recordings of Indian doctors and medical students using mostly synthetic speech. For speakers the model had not heard before, this fine-tuning resulted in a relative improvement from a 22.44% error rate to a 12.17% error rate. The model ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English. It posts the lowest error rate on David AI's 21-language diarization leaderboard and ranks first in VoiceArena's 14-system comparison for speaker attribution.

Agora used word-level timestamps from Nemotron 3.5 Transcribe to show transcripts alongside video replays in a classroom demo.