NVIDIA Open-Sources Nemotron 3.5 Transcribe: One Promptable Model for Transcription in 40 Languages, Speaker Labels, Translation and Audio Q&A, Plus 460,000 Hours of Granary v2 Data

One recording, a different output for each prompt

NVIDIA released Nemotron 3.5 Transcribe on October 28, with weights on Hugging Face and ModelScope. Its selling point is that it is promptable: give it a meeting recording and "transcribe this and label each speaker" returns a speaker-attributed transcript, "translate this into English" returns a translation, and "what decisions were made in this meeting?" returns an answer. NVIDIA's stated scope: transcription in 40 languages, translation to or from English for 31 of them, and speaker attribution in 13; recordings up to an hour long with up to eight speakers. Audio question answering is trained in English only.

For unfamiliar terms, the prompt can list candidates, such as "HbA1c, metformin" in a medical setting, with an instruction to use them only when the audio supports them.

Parameters, license and results

The model card lists 34B total parameters with about 4.3B active per token, including an audio encoder of roughly 800M parameters; the language side is a hybrid Mamba–Transformer mixture of experts. It ships in BF16, FP8 and NVFP4 under the OpenMDW-1.1 license and is cleared for commercial use. NVIDIA's reported results: first among open-weight models on the Hugging Face Open ASR English leaderboard at 5.33% average word error rate (the chart excludes API models); 2.8% on Artificial Analysis's AA-WER, tied for first among open-weight models with a model that has 24B active parameters; and an 18.26% diarization error rate on David AI's 21-language leaderboard, the lowest of the 14 systems tested.

The fine-tuning data ships too

The Granary v2 manifest released alongside it covers 463,062 hours from 16 source corpora in 48 languages, with verbatim, normalized and disfluency-free transcripts for each recording. In a demo on a benchmark of Indian doctors' speech, a better prompt alone cut the error rate from 30.8% to 23.8%; about 14 GPU-hours of fine-tuning then took held-out speakers from 22.44% to 12.17%.

Where it fits

The model card is explicit that this is an offline model and that its internal chunking does not make it a real-time streaming service. It suits recorded meetings, lecture replays and medical dictation rather than low-latency voice agents. NVIDIA also advises spot-checking names, speaker changes and answers against the original recording.