NVIDIA released Nemotron 3.5 Transcribe on 29 October, an open-weight model that turns a recording into a transcript, a translation, a speaker-labelled script or an answer about what was said, depending on the instruction you give it. NVIDIA described the release on the Hugging Face blog, and the weights are on Hugging Face and ModelScope, so a team can host the model itself.

The model transcribes 40 languages and can translate into or out of English for 31 of those. It labels speakers in 13 languages, accepts recordings of an hour or less, and keeps track of up to eight voices. Answering questions about audio is trained in English only. It works on finished recordings and does not stream, which rules it out for captioning a call as it happens.

Control comes through plain-language prompts. NVIDIA’s example for a clinic asks for a transcript with a list of possible terms such as HbA1c and metformin, to be used only when the audio supports them. That last condition is aimed at a familiar failure, where a model that has been told to expect a drug name writes it down whether or not anyone said it.

On Hugging Face’s Open ASR ranking, NVIDIA says the model tops the open-weight field for English, with a 5.33 percent average word error rate. Artificial Analysis, which runs its own speech-to-text ranking, has it sharing the top open-weight spot at 2.8 percent. The model it ties has 24 billion active parameters. Transcribe activates about 4 billion of its 30 billion, which is where the efficiency story comes from.

The speaker-labelling results deserve a closer look. On David AI’s diarization ranking across 21 languages (diarization means working out who spoke when), Transcribe posts an 18.26 percent error rate. The runner-up at 18.38 percent is NVIDIA’s own Nemotron 3 Diarization. VoiceArena’s comparison of 14 systems has the same shape: Transcribe at 32.4 percent, NVIDIA’s Multitalker Parakeet pairing at 34.2 percent. NVIDIA holds the top two places on both boards, and the gaps on the first are small.

Speed numbers come with a catch. NVIDIA reports 100 hours of audio processed in less than three minutes on one Blackwell-generation RTX PRO 6000 server GPU, but that run used a compressed NVFP4 version of the model that has not shipped. The quantized releases and the local runtime are both listed as coming soon. What you can serve today is the standard checkpoint through vLLM.

The more durable part may be Granary v2, the training-data pipeline. Its public manifests list 463,062 hours of speech in 48 languages, drawn from 16 source corpora. The first version covered 25 languages. Audio has to be downloaded separately under each corpus’s own license, and NVIDIA says those totals describe the published data, not the full mix used to train Transcribe.

NVIDIA also ran a medical case study on a benchmark from Eka Care made of recordings by doctors and medical students in India. Telling the model in the prompt to write English in Latin script cut errors on the full benchmark from 30.8 percent down to 23.8 percent. Fine-tuning on synthetic and real recordings, roughly 14 hours of one GB300 GPU, then took held-out speakers from 22.44 to 12.17 percent, a 45.8 percent relative drop. That beat an open medical-specialist model at 13.79 percent and a general-purpose model at 14.92 percent. The specialist still wins on brand-name drugs, and error on general English speech crept up, from 1.54 percent to 1.69 percent.

Most of these results are NVIDIA’s own measurements, with Artificial Analysis the one independent exception. A team paying per minute for a hosted transcription service can now test an open model on its own audio, and NVIDIA’s advice of 100 or more real clips with labels is a sensible size for that trial.

Reported by NVIDIA on the Hugging Face blog on 29 October 2026.