NVIDIA Shows Nemotron 3.5 ASR Fine-Tuning Cut Saudi Arabic Dialect WER to 29.96%

NVIDIA has published a fine-tuning workflow showing that its multilingual Nemotron 3.5 ASR model can be substantially improved for Saudi Najdi and Hijazi speech, cutting word error rate on the target test split from 55.05% to 29.96%. The result is a reminder that support for a language label such as Arabic does not guarantee reliable performance on local dialects, recording conditions, or deployment-specific speech.