Skip to content

NVIDIA Announces Fine-Tuning of Nemotron 3.5 ASR for Saudi Arabic Dialects

NVIDIA Nemotron 3.5 ASR supports multilingual streaming transcription and is now fine-tuned for Saudi Arabic dialects, including Najdi and Hijazi.

Sources in the coveragePerformance and quality statements from a vendor are claims unless the coverage cites independent measurements.

Evidence & verification 3 statements

Specifications, source statements and editorial judgement have different scopes. A citation is not an independent performance test.

Word Error Rate (WER) reductionSource statement

from 55.05% to 29.96%

developer.nvidia.com ↗Source observation date not recorded.
English WER improvementSource statement

from 11.04% to 10.42%

developer.nvidia.com ↗Source observation date not recorded.
Trainable parametersSource statement

230.4 million

developer.nvidia.com ↗Source observation date not recorded.

NVIDIA announced the fine-tuning of their Nemotron 3.5 ASR model for Saudi Arabic dialects, specifically Najdi and Hijazi. This fine-tuning process uses the NVIDIA NeMo framework and reduces the word error rate (WER) from 55.05% to 29.96% on the target test split.

Key facts

Fact Detail The source says
Word Error Rate (WER) reduction from 55.05% to 29.96% “Fine-tuning on 133.7 hours of Najdi and Hijazi speech reduced word error rate from 55.05% to 29.96% on the target test split”
English WER improvement from 11.04% to 10.42% “improved English performance from 11.04% to 10.42% WER without degrading other Arabic dialects”
Trainable parameters 230.4 million “Updating all 24 encoder layers achieved the lowest error rates at the cost of 230.4 million trainable parameters”

What happened

NVIDIA has announced the fine-tuning of their Nemotron 3.5 ASR model for Saudi Arabic dialects, specifically Najdi and Hijazi. The fine-tuning process uses the NVIDIA NeMo framework and reduces the word error rate (WER) from 55.05% to 29.96% on the target test split. This workflow involves curating a low-resource corpus, building a weighted replay mix, fine-tuning with efficient batching, and evaluating transcription quality on an independent set. The process also improves English performance from 11.04% to 10.42% WER without degrading other Arabic dialects. Updating all 24 encoder layers achieved the lowest error rates but at the cost of 230.4 million trainable parameters. Freezing layers reduces compute needs when data or memory is limited.


Source: NVIDIA

Keep following technology

Continue from here.

Explore more coverage or subscribe to Fathom Layer updates.

Back to the news desk ↗

Occasional verified technology updates. Unsubscribe anytime. See our privacy policy.