Nvidia uses Saudi voice data to sharpen dialect recognition

Nvidia says it fine-tuned its Nemotron 3.5 ASR speech-to-text system using the Saudi Audio Dataset for Arabic, or SADA, released by the Saudi Data and AI Authority with the Saudi Broadcasting Authority. The work focused on Najdi and Hijazi, a useful reminder that an AI system can handle Arabic broadly while still struggling with the vocabulary, pronunciation, and recording conditions of regional speech. Using 133.7 hours of selected recordings, Nvidia reported that its word-error rate on a Saudi-dialect test set fell from 55.05% to 29.96%. The result is not a new consumer product, but a demonstration of how locally representative data can make live captions, call-center transcription, and voice assistants more useful for Saudi speakers while preserving a model’s wider multilingual ability.