Nvidia’s new speech model is open. The error rate is Nvidia’s

Nvidia posted Nemotron 3.5 Transcribe on 28 October and called it one model for a job that usually takes several. A recording goes in. The model can return a verbatim transcript, a cleaned written version, a translation, a label for who spoke, or an answer about what was said, depending on the instruction. It covers 40 languages, recordings up to an hour, and word-level timestamps tied to a speaker. The language model underneath is Nemotron 3.5 Lightning, a 30-billion-parameter mixture of experts, with about 4.3 billion parameters active on a given pass, plus a speech encoder and Nvidia’s existing diarization model. One pass is the product claim. It is not a claim that the transcript is finished enough to file.

The accuracy number is Nvidia’s, on public sets. On the nine English tests of the Open ASR leaderboard, the company says the mean word-error rate is 4.71 percent. On the multilingual board it says 4.62 percent. Nvidia’s blog adds that this is first among open models on the Hugging Face Open ASR list, and second among open-weight models on Artificial Analysis, with about 80 percent fewer active parameters than the open model ahead of it there. A word-error rate is the share of words wrong. It is not a measure of whether a name, a part number or a safety instruction survived. The company also says, on its own tests, that the model beats the usual chain of a transcriber plus a separate speaker system when people talk over each other. That comparison is Nvidia’s.

The part a buyer can check is the download. The weights are open, in BF16, FP8 and NVFP4, on Hugging Face. Nvidia says they run on RTX Spark and DGX Spark machines, on Jetson AGX Thor, and on a GeForce RTX 5090, through NeMo-Speech.cpp or vLLM. Open weights mean a factory can keep the audio on its own machine. They do not mean the model is free of the data it was trained on, or that an hour-long recording of a noisy line will match the leaderboard. Nvidia did not publish a price, because there is no API in the announcement. The cost is the computer it runs on.

For a robotics team the relevant use is narrow. A model that can say who said what, on a recording that never leaves the building, is a way to label demonstrations and incident audio without sending them out. It is not a controller. It does not move a joint, and the 4.71 percent was not measured on a factory floor. Until someone publishes an error rate on that kind of audio, Transcribe is a strong open transcriber with a number attached by its maker.