whisper-large-v3-turbo-arabic-dialectal-v2

whisper-large-v3-turbo-arabic-dialectal-v2

v2 — openai/whisper-large-v3-turbo (809M) fine-tuned for multi-dialect Arabic on the dialect-balanced dataset oddadmix/lahgtna-v3-small (52k train / 2.6k test, seed 42, undiacritized output).

Open weights. This is v2, trained on the balanced, leakage-reduced data. v1 (trained on the earlier augmentation-expanded set, with the full cross-model comparison + bundled fine-tuning scripts) is at oddadmix/whisper-large-v3-turbo-arabic-dialectal.

Results (balanced test split ~2,600 clips, clean_text WER/CER)

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 openai/whisper-large-v3-turbo untouched; fine-tuned (v2) is this model. Text is clean_text-normalized (undiacritized), matching the WER scoring. Press play to listen.

Full fine-tuning code, dependencies, and lessons are in the v1 repo (train.py, normalize.py, FINETUNE.md). This v2 uses the same recipe on the balanced dataset.