Nvidia Nemotron 3.5 Transcribe: one open speech model

Nvidia released Nemotron 3.5 Transcribe on October 28, 2026, an open-weight speech model that does four jobs in one. It transcribes audio, labels who is speaking, translates and answers questions about a recording. Developers steer it with plain-language prompts. That could replace a pipeline of three or four separate speech tools with a single download.

Nvidia also published Granary v2, the dataset it trained on, and the tools it used to build it.

What the model does

According to Nvidia's announcement on the Hugging Face blog, one checkpoint handles transcription, speaker labeling, translation and audio Q&A. Speaker labeling, often called diarization, means tagging each line with who said it.

Capability	Coverage, per Nvidia
Transcription	40 languages
Translation	31 languages, to or from English
Speaker attribution	13 languages, up to 8 speakers
Recording length	Up to 1 hour
Context window	64k tokens

Prompts change the output style. Nvidia's example turns spoken words like "twenty-five euros" into "€25" when asked for written form.

There is one clear limit. "Transcribe is built for recorded audio, not live streaming," the blog says. Quantized and local versions, which would run on smaller hardware, are still to come.

How it is built

The model is a mixture of experts (MoE). In an MoE model, only a few parts of the network run for each piece of input, which keeps it fast. It mixes Mamba layers, a newer design that handles long sequences cheaply, with standard Transformer layers.

The blog gives 30 billion total parameters, with about 4 billion active per token. The model card counts the audio encoder too. It lists 34 billion total and about 4.3 billion active, including an audio encoder of about 800 million parameters.

The license is OpenMDW-1.1. The card says the model "is ready for commercial or non-commercial use." Nvidia used the same license family for Nemotron 3.5 Super VL, its 120B open vision model, earlier this month.

How it scores

Most of the benchmark numbers come from Nvidia itself. One independent ranking backs up the main claim.

Artificial Analysis, which tests speech models across providers, ranks Nemotron 3.5 Transcribe second among open-weight models. Its AA-WER is 2.8%. WER is word error rate, the share of words a model gets wrong, so lower is better. The same page lists a 3.4% WER in its main table without explaining the gap. It prices the model at $2.00 per 1,000 minutes through the provider DeepInfra.

Nvidia's own figures include these:

• Multilingual: a mean WER of 4.62% across seven languages on Hugging Face's Open ASR tests.
• Long recordings: a mean WER of 10.98% on long-form audio.
• Conversations: 1.8% WER on AA-AgentTalk, against 2.3% for the next-best open model.
• Speed: 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell GPU, using a forthcoming NVFP4 version.

That speed figure needs care. The model card says those serving optimizations are not yet in the open-source release. It expects the open NeMo and vLLM setup to run roughly twice as slow.

Granary v2 and custom tuning

Granary v2 holds 463,062 hours of audio in 48 languages, drawn from 16 source corpora. The first Granary covered 25 languages. Nvidia also shares the NeMo Curator pipeline it used, so teams can run the same cleanup on their own recordings. The audio itself must be downloaded separately under each corpus's original license.

Nvidia tested fine-tuning on medical speech with Eka Care data. Error rates on held-out speakers fell from 22.44% to 12.17%. General English got slightly worse: LibriSpeech test-clean WER rose from 1.54% to 1.69%. Nvidia recommends at least 100 labeled clips for a custom tune.

What this means for developers

An open model that transcribes, labels speakers and translates in one pass changes the build-versus-buy math for speech features.

• Test it against your current stack. If you chain a transcription model with a separate diarization tool, one model may cut latency and glue code. Compare on your own audio, not only on leaderboards.
• Keep it away from live audio. Nvidia says it is built for recordings. Meeting notes, podcasts and call archives fit; live captions do not, yet.
• Plan the hardware. About 4 billion active parameters makes it cheaper to run than its size suggests, but the full weights still need a data center GPU until smaller builds ship.
• Fine-tune for jargon. The medical test shows big gains from a small, focused dataset. Legal, medical and support teams with their own vocabulary should budget for a tune and check general accuracy afterward.
• Check the output. Nvidia's own advice is to "spot-check names, speaker changes, and answers against the recording." Build a review step into any workflow where a wrong name or speaker matters.

The real test is whether developers trust one model with four jobs. Open weights let them find out on their own data, which is more than any leaderboard can show.