Audio Insight (Ar& En)

Developed Audio Insight, an intelligent NLP web assistant that supports Arabic dialects (including Egyptian colloquial) and English, feature:

•

Audio Transcription: Fine-tuned a small version of OpenAI Whisper on Arabic and English speech using the fineweb_2 dataset.

•

Achieved a final WER of 7%, representing a 55% improvement over the pre-trained baseline.

•

Summarization: Fine-tuned Googleʼs mT5 to summarize bilingual text and transcriptions with a ROUGE score ≥ 0.7.

•

Fine-tuned Qwen1.5-Instruct using LoRA via LLaMA Factory with knowledge distillation from Mistral, and implemented a real-time cost

estimation module to optimize instruction-based NLP performance in low-resource settings.

•

Title Generation: Used a chat template instruction model to generate context-aware, concise titles from input summaries.

•

Backend: Developed REST APIs for all features using Flask, and deployed components in Docker containers for scalability and modularity.

•

Guide Bot: Created an interactive guide bot to assist users with accessing and understanding the platform's features