qwen3-asr-1.7b-arabic-dialectal-v2

v2 — Qwen/Qwen3-ASR-1.7B-hf (1.7B) fine-tuned for multi-dialect Arabic on the dialect-balanced dataset oddadmix/lahgtna-v3-small (seed 42, undiacritized). Trained off-machine on a larger GPU.

Open weights. Part of the v2 dialectal-ASR family.

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 — fine-tuned output (one clip per dialect)

13 held-out test clips, one per dialect, transcribed by fine-tuned (v3) (this model). Text is clean_text-normalized (undiacritized), matching the WER scoring. Press play to listen.

Base-column transcripts (Qwen/Qwen3-ASR-1.7B-hf) omitted — base weights weren't available locally.