A 30 September NVIDIA technical blog describes adapting the Nemotron 3.5 automatic speech-recognition model to the Najdi and Hijazi dialects used in Saudi Arabia. The team used 133.7 hours of target-dialect speech from the Saudi Audio Dataset for Arabic (SADA). The blog says the curation retained 103,559 utterances from an initial 125,490, or 82.5%. This is the company’s own training and evaluation experiment, not an independent, nationwide accuracy assessment.

NVIDIA reports that word error rate (WER) on its Najdi-Hijazi target test split fell from 55.05% to 29.96%. WER across the full SADA test set declined from 58.84% to 35.61%. WER measures substitutions, deletions and insertions relative to the number of words; it is not equivalent to saying that the model became ‘25% more accurate.’ In the same experiment, FLEURS English WER changed from 11.04% to 10.42%, and Arabic WER from 12.67% to 11.41%. These figures belong to separate test sets and should not be conflated.

The training recipe mixes target-dialect speech with weighted English and Arabic examples from FLEURS and groups utterances of similar duration to manage training computation. The best reported SADA result followed 12,000 training steps and about 4.5 hours on two GPUs. Those are NVIDIA’s experimental conditions; they do not guarantee the same cost or result with different data, model versions or infrastructure.

The post stresses that results vary by evaluation set and that a gain on one test cannot be generalized automatically to other dialects or real-world settings. Curation choices, background noise, microphones, speaking rate and transcript conventions can all affect WER. NVIDIA also reports that offline decoding options lowered WER while adding latency; a wider attention context can add about 800 milliseconds. Live captioning, call-archive transcription and meeting records therefore involve different trade-offs.

Speech recognition adapted to Saudi dialects could be a technical component for customer service, media archives or voice interfaces. The source does not say that this model has been deployed by Saudi businesses or won a commercial contract. Its measurements show how the output changed on the specified test sets under the published recipe. Real deployments would still need local user testing, error analysis, privacy and consent controls, accessibility review, operating-cost assessment and appropriate human oversight.

The table reports WER percentages and keeps test sets separate. Before-and-after values are the measurements in NVIDIA’s technical post. They are experimental results and do not claim to represent every dialect or speaker in the country. The supported conclusion is that dialect-targeted fine-tuning improved the measured test performance in this experiment; performance in actual use requires separate validation.