Lafz2Lafz: An End-to-End AI Pipeline for Urdu OCR, English Translation, Summarization, and Text-to-Speech Synthesis

Abstract:

A large amount  of Urdu literature and educational material is available  in printed and image-based formats, making it hard to search, analyze, and access digitally. This research introduces an integrated platform, Lafz2Lafz (Every Word Counts), that uses Optical Character Recognition (OCR) to extract Urdu text from images, translates it into English, generates short summaries and converts the output into audio through an automated pipeline.

The main part of this work is the detailed comparison of multiple advanced models at each processing stage. For OCR, we tested Kraken and TrOCR on same datasets of scanned Urdu textbooks. For translation, we compared three approaches MBERT, NLLB (No Language Left Behind), and Google Translator. For summarization, we compared two abstractive methods, DistilBART and T5. For voice output we use Kokoro TTS. Through this evidence-based selection process, Kraken was chosen for OCR due to its amazing performance with Urdu Nastalique script (lower CER and better handling of cursive ligatures), while NLLB was selected for translation based on its strong vocabulary , reliable handling of domain-specific terminology, and self-contained deployment without external API dependencies. For summarization DistilBART was chosen due to its ROUGE score and lower processing latency.

The selected models were combined into a unified web-based platform with a Flask backend, letting users to upload Urdu images and receive extracted text, English translations, summaries and voice output through a simple interface. This remove the need for multiple separate tools and manual data transfer between processing stages, really improving workflow efficiency and data integrity.