nemotron-3.5-asr-arabic-dialectal-v2

nemotron-3.5-asr-arabic-dialectal-v2

v2 — nvidia/nemotron-3.5-asr-streaming-0.6b (cache-aware streaming RNNT, native Arabic) fine-tuned on the dialect-balanced oddadmix/lahgtna-v3-small (52k/2.6k, seed 42).

Open weights. v1 at oddadmix/nemotron-3.5-asr-arabic-dialectal.

Results (v3 2,445-clip balanced test)

Load with NeMo main (ASRModel.from_pretrained). Fine-tuning code + FINETUNE.md in the v1 repo.

<!-- v2-comparison-examples -->

---

Per-dialect accuracy (v3 balanced test, clean_text)

WER / CER on each dialect's 200 held-out clips, easiest to hardest. Gulf/Levantine transcribe best; Maghrebi (Tunisian/Algerian/Moroccan) is hardest — consistent across every model.

v2 comparison — all models on the balanced dataset

Same dialect-balanced test split (~2,600 clips, oddadmix/lahgtna-v3-small), same clean_text scoring, seed 42.

whisper-large-v3-turbo-v2 is the best; Whisper fine-tunes furthest. nemotron is the streaming option. (v1 models were trained on the earlier augmentation-expanded set.)

Examples — base vs. fine-tuned (one clip per dialect)

13 held-out test clips, one per dialect. Base is nvidia/nemotron-3.5-asr-streaming-0.6b untouched; fine-tuned (v2) is this model. Text is clean_text-normalized (undiacritized), matching the WER scoring. Press play to listen.