Speech tooling usually means stitching together a transcriber, a translator and a diarizer, then hoping the handoffs hold. Nvidia’s Nemotron 3.5 Transcribe collapses that stack into a single open-weight system that accepts written commands. Forty tongues get transcribed, 31 of those also get translated against English, and 13 come with speaker labels. An hour-long file with up to eight distinct voices fits inside the limits.
The efficiency is the story. Weights total 30B parameters, yet only around 4B switch on for each step, so one GPU is enough, whether that is a compact DGX Spark or a data-center card. On a Blackwell-generation RTX PRO 6000 server card at NVFP4 precision, Nvidia puts throughput at 100 hours of audio inside three minutes.
Independent tests support the claim. Hugging Face’s English Open ASR Leaderboard places it atop open-weight entrants at a 5.33 percent average word error rate, and Artificial Analysis recorded 2.81 percent, nearly matching Voxtral Small at 2.77 percent despite activating roughly a sixth as many parameters. VoiceArena put it first on conversations where people interrupt, losing 8.32 percent of words against a typical mark near 31 percent.
Alongside the model Nvidia is opening Granary v2, roughly 844,000 hours of speech spanning 48 languages, up from 643,000 hours across 25. A tuning experiment on audio from Indian physicians lowered word error from 30.8 to 23.8 percent with better prompting and to 11.8 percent after adaptation, achieved in about 14 GPU hours on a GB300.
Caveats apply. It is built for recordings rather than live streams, and Nvidia recommends spot-checking names and speaker turns, since the failure it most wants to avoid is a confident attribution to the wrong voice.
