NVIDIA reports that fine-tuning on 133.7 hours cut word error rate from 55.05% to 29.96% on the target test split.

NVIDIA fine-tuned Nemotron 3.5 ASR on 133.7 hours of Saudi Najdi and Hijazi speech using the NVIDIA NeMo framework. The recipe combined weighted replay mixing with FLEURS data, duration-based bucketing, and partial encoder unfreezing. Word error rate on the target test split fell from 55.05% to 29.96%.

NVIDIA developer blog authors Imane Khaouja, Amine El Khair, Meshari Alaeena, Zahra Al-Kaf, and Abdulrahman Alkhamees described the work. Nemotron 3.5 ASR supports multilingual streaming transcription across 40 language-locales, but deployment-specific dialects such as Saudi Najdi and Hijazi benefit from targeted fine-tuning.

Builders can adapt a multilingual ASR model to a low-resource dialect while preserving existing language performance. Training all 24 encoder layers produced the lowest error rates but required 230.4 million trainable parameters. Operators can also switch to a larger attention context of 13 lookahead frames and beam-8 MALSD decoding to lower word error rate by 2.71 absolute points without retraining.

NVIDIA Nemotron 3 Diarization extends the workflow to speaker-attributed transcription for up to eight speakers. The diarization model aligns speaker boundaries with ASR timestamps for multi-speaker environments. NVIDIA offers an ASR fine-tuning notebook and a Nemotron ASR fine-tuning skill for guided implementation. A Hugging Face blog describes how to add speaker diarization to the fine-tuned ASR pipeline.

What matters

• NVIDIA fine-tuned Nemotron 3.5 ASR on 133.7 hours of Saudi Najdi and Hijazi speech.
• The recipe cut target WER to 29.96% and lowered English WER to 10.42% without degrading other Arabic dialects.
• NVIDIA Nemotron 3 Diarization extends the recipe to speaker-attributed transcription for eight speakers.

Why it matters

NVIDIA Nemotron 3 Diarization extends the recipe to speaker-attributed transcription for eight speakers.

This GenAI News article was prepared in original wording using reporting and materials published by NVIDIA Developer Blog. Source reference: https://developer.nvidia.com/blog/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages/.

Drafted by the GenAI News review pipeline.