NVIDIA details fine-tuning Nemotron ASR for Saudi dialects

NVIDIA has published a technical walkthrough on fine-tuning Nemotron 3.5 ASR for Saudi Arabic dialects. The company says the same method can be applied to other low-resource languages and domains.

Nemotron 3.5 ASR supports multilingual streaming transcription across 40 language-locales, including Arabic. NVIDIA acknowledges that it can still struggle with regional speech, such as Najdi and Hijazi, and with local recording conditions. Fine-tuning only on the target dialect risks degrading the model's other languages. According to a summary of the post's results, Saudi-dialect word error rate (WER) fell from about 55% to 30% with no regression on English.

The recipe uses the NeMo framework, the SADA 2022 and FLEURS datasets, and Cache-Aware FastConformer-RNNT as the model. It combines:

• Minimal curation. NVIDIA filtered out clips with unusable labels and probable misalignment, including SADA's "inaudible" annotations, which the model cannot output. This kept 103,559 of 125,490 utterances (133.7 hours, 82.5% of the set).
• Corpus-specific quality thresholds in NeMo Curator. A default UTMOS cutoff of 3.0 would have rejected almost all the data, so the team used UTMOS ≥ 1.25 and SIGMOS ≥ 1.5. About 85% of duration-valid samples passed.
• Replay mixing to limit catastrophic forgetting, partial encoder unfreezing, and length bucketing for streaming encoders.

The baseline run took 12,000 steps on two RTX PRO 6000 Blackwell Workstation GPUs.

NVIDIA says the techniques are not universal defaults. Replay only protects the languages present in its data, and partial unfreezing needs re-tuning whenever the data mix changes. The company also says the results are not evidence of performance on every Arabic dialect or deployment environment.