vadimbelsky/qwen3-asr-arabic-ksa

Qwen3-ASR Arabic — KSA Saudi Dialect

Fine-tuned Qwen/Qwen3-ASR-1.7B for KSA Saudi Arabic dialect speech recognition.

Sequential fine-tuning: Base Qwen3-ASR -> UAE model -> this KSA model.

Results

Metric	Zero-shot (base)	Fine-tuned	Improvement
WER	14.41%	11.49%	-20%
CER	6.57%	5.78%	-12%

Evaluated on 849 KSA Arabic validation samples.

What improved

• Better handling of KSA dialect expressions and vocabulary
• Removes spurious punctuation that the base model adds
• Matches informal Saudi dialect spelling conventions

Training Details

• Base model: vadimbelsky/qwen3-asr-arabic-uae (itself fine-tuned from Qwen3-ASR-1.7B on UAE data)
• Training data: ~7,350 KSA Saudi Arabic dialect samples from vadimbelsky/KSA_Arabic_English_Dataset_13k
• Strategy: Audio encoder frozen, only LLM decoder fine-tuned (84.4% of params)
• Precision: bfloat16
• Epochs: 3
• Effective batch size: 32 (batch 2 x gradient accumulation 16)
• Learning rate: 1e-5 with linear schedule (lower than UAE stage)
• Gradient checkpointing: enabled
• Text normalization: Diacritics removed, alef/teh marbuta normalized, punctuation stripped

Usage

from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained("vadimbelsky/qwen3-asr-arabic-ksa")
result = model.transcribe("audio.wav", language="Arabic")
print(result)

Or with transformers directly:

from transformers import AutoModelForCausalLM, AutoProcessor

model = AutoModelForCausalLM.from_pretrained("vadimbelsky/qwen3-asr-arabic-ksa")
processor = AutoProcessor.from_pretrained("vadimbelsky/qwen3-asr-arabic-ksa")

Limitations

• Trained on synthetic/generated Arabic speech data
• Optimized for KSA Saudi dialect — may not generalize to other Arabic dialects
• Numbers are sometimes transcribed differently (e.g. digits vs spelled out)
• Short utterances only (training data mostly < 20s)

Related Models

• vadimbelsky/qwen3-asr-arabic-uae — UAE Emirati dialect (9.98% WER)

License

Apache 2.0 (same as base model)