Nvidia launches a promptable transcription model

Nvidia launched Nemotron 3.5 Transcribe on Wednesday, and the company is leaning hard into a newer idea: transcription that takes instructions before it listens. The open-weight model is meant for multilingual speech-to-text, but it also lets users steer the output with prompts, from domain terms to speaker labels to formatting rules for numbers.

That matters because transcription has usually been a blunt tool. You feed it audio, you get text back, and then you clean up the mess. Nvidia is trying to fold more of that cleanup into the model itself. It can produce a verbatim transcript or a cleaned-up version, answer questions about the audio, and bundle translation into the same checkpoint. For developers, that could mean fewer separate pieces to wire together.

The numbers are strong, at least by Nvidia’s telling. The model ranks first on Hugging Face’s Open ASR open model leaderboard for English, with an average word error rate of 5.33%, and it tops two speaker-attribution benchmarks as well. On Artificial Analysis’s benchmark, it posts a 2.81% word error rate, just behind Voxtral Small’s 2.77%, while using about 4 billion active parameters versus 24 billion. The comparison is about active parameters, though, not runtime or memory use.

Nemotron 3.5 Transcribe is a 30B mixture-of-experts model with 4 billion active parameters. It handles 40 languages, translates 31 of them to or from English, and can distinguish up to eight speakers. Speaker attribution works for 13 languages, and audio Q&A was trained in English. It supports recordings up to an hour long, but not live streaming.

Nvidia says the model can run on a single GPU, with support ranging from Jetson AGX Thor and DGX Spark to Blackwell data-center GPUs. Because it’s released under the OpenMDW-1.1 license, users can fine-tune it. In one medical transcription case study, Nvidia says fine-tuning cut word error rate on held-out speakers from 22.44% to 12.17%, and the adaptation workflow took about 14 GPU-hours on a single GB300. Nvidia is also releasing Granary v2, a data curation pipeline with public manifests covering roughly 463,062 hours across 48 languages and 16 source corpora.