Nvidia’s open transcription model comes with the recipe to adapt it

Nemotron 3.5 Transcribe is an open-weight model that handles transcription in 40 languages, translation for 31 of them to or from English, and speaker labeling in 13, for recordings up to an hour with up to eight speakers. It takes plain-language instructions, so the same meeting file can produce a transcript, a translation or an answer about what was decided. It pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model and is built for recorded audio rather than live streaming. The checkpoint is on Hugging Face and ModelScope, with vLLM and SGLang-Omni support and hosted partner endpoints, and Nvidia says quantized versions are coming for local hardware such as RTX Spark laptops (RTX Spark update).

Plenty of speech models top a leaderboard. What sets this release apart is Granary v2, a public manifest covering about 463,000 hours across 48 languages, released with the curation pipeline and fine-tuning recipes. In Nvidia’s medical case study, simply naming the language and script in the prompt cut word error rate from 30.8% to 23.8%; fine-tuning took it to 11.8% on the full set, and on held-out speakers fine-tuning cut error by 45.8% relative to the best prompt. The adaptation took about 14 GPU-hours on one GB300, within reach for a hospital system or call center, not only a lab.

Efficiency helps too. Only about 4 billion of 30 billion parameters are active per token, and Nvidia says it ties for first among open-weight models on Artificial Analysis’s index against a model with roughly 24 billion active. Open weights and on-prem deployment mean sensitive recordings can stay in the building. It extends Nvidia’s earlier open diarization model that tracks eight speakers (Nemotron 3 Diarization) and competes with smaller permissive options like IBM’s Granite Speech (Granite Speech 5.0).

Most leaderboard claims come from Nvidia’s own blog, though several benchmarks are run by others. The medical study leaned heavily on synthetic speech, and Nvidia notes a specialist model still beats it on exact brand-name drugs. In medicine a misheard drug name is a potential harm, not a typo, so Nvidia’s advice to spot-check names, speaker changes and answers should be the rule anywhere a transcript becomes a record.

There is a labor angle as well. Cheap, accurate open transcription with speaker labels pushes more work away from human transcriptionists and note-takers, and easy who-said-what at scale makes recording and analyzing conversations much cheaper, which raises consent questions in workplaces and classrooms. More in our Nvidia coverage and AI coverage.