ARABIC (EGYPTIAN) SPEECH TO TEXT

Fine-tuned OpenAI’s Whisper-small (244M parameters) on 20,000 Mozilla Common Voice audio recordings for dialectal Egyptian Arabic. Dialectal speech breaks vanilla models due to colloquial phonetic shifts; this implementation engineered a robust audio processing pipeline featuring 16kHz resampling, log-mel spectrogram extraction, Arabic diacritic removal, and FP16 PyTorch training on RTX 3060 with Gradio streaming inference.