Model Releases
Fine-Tuning NVIDIA Nemotron for Saudi Arabic Dialects, with a Path to Other Languages
NVIDIA Nemotron 3.5 ASR was fine‑tuned on 133.7 h of Saudi Najdi and Hijazi speech using weighted replay with FLEURS data, duration bucketing, and selective encoder unfreezing; this reduced the target
NVIDIA Nemotron 3.5 ASR was fine‑tuned on 133.7 h of Saudi Najdi and Hijazi speech using weighted replay with FLEURS data, duration bucketing, and selective encoder unfreezing; this reduced the target dialect WER from 55.05 % to 29.96 % while slightly improving English performance (11.04 %→10.42 %) without harming other Arabic dialects. Updating all 24 encoder layers yielded the best error rates at a cost of 230.4 M extra parameters, whereas freezing layers can lower compute needs; additionally, increasing attention context to 13 look‑ahead frames and applying beam‑8 MALSD decoding achieved an absolute WER drop of 2.71 % with ≈800 ms added latency suitable for batch transcription. The workflow also supports speaker‑attributed transcription up to eight speakers via Nemotron 3 diarization, and users can reproduce the adaptation through the provided notebook or guided skill on Hugging Face.
Related
- Evaluate Clinical ASR Models Faster with Agent Skills and NVIDIA Nemotron Speech
- How to Fine-Tune Nemotron 3.5 ASR for Your Language, Domain, or Accent
- nvidia/NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4 · Hugging Face
Source: NVIDIA Developer | 2026-10-01