A Saudi speech dataset helped one of NVIDIA’s speech recognition models understand two Saudi dialects much better, according to a technical study NVIDIA published on its developer blog on 30 September 2026.
The team fine-tuned Nemotron 3.5 ASR, a model that transcribes speech in 40 language-locales, using 133.7 hours of Najdi and Hijazi speech from SADA, a Saudi Arabic audio dataset. On those two dialects, the share of words transcribed wrongly fell from 55.05% to 29.96%. Across the full SADA test set, which also includes other dialects, it fell from 58.84% to 35.61%. Accuracy on English held and even improved slightly, with word errors falling from 11.04% to 10.42%, and the training took about 4.5 hours on two GPUs.
Arab News, citing the Saudi Press Agency, reported on 3 October that SADA was launched by the Saudi Data and AI Authority (SDAIA) with the Saudi Broadcasting Authority, and that most of its recordings come from Saudi broadcast programmes.
Why it matters: before the adaptation, the model got more than half the words wrong in these dialects. The uses NVIDIA names, such as voice agents, live captions and contact centres, only work for Saudi users if the system understands how they actually speak. The study also shows how an existing model can be adapted with a public local dataset in hours, rather than a new model being built from scratch.
What to watch: NVIDIA cautions that the workflow “doesn’t generalize into evidence for every Arabic dialect or deployment environment”. Only Najdi and Hijazi speech was used in training.