Automating Ledger Digitization: Performance Evaluation of On-Device OCR Models for Recognizing Handwritten Financial Scripts

The digitization of handwritten financial records, such as ledgers, cash books, and account registers, represents a significant challenge in the domain of intelligent document processing and financial technology. Traditional Optical Character Recognition (OCR) engines, while effective for printed text, frequently fail to accurately interpret cursive, regional-language, or domain-specific handwritten financial scripts. This research investigates and benchmarks the performance of three on-device OCR models — Google ML Kit Text Recognition v2, Apple Vision Framework, and Tesseract OCR v5 — for recognizing handwritten financial text under real-world conditions. A curated dataset of 1,200 handwritten ledger images sourced from small and medium enterprises across India, encompassing Hindi, English, and mixed-script entries, is used for evaluation. Performance is assessed using Character Error Rate (CER), Word Error Rate (WER), and token-level F1-Score. Results demonstrate that fine-tuned Google ML Kit achieves the best CER of 6.3%, with further improvement to 4.5% after applying a post-processing language correction layer. The findings carry direct implications for FinTech platforms such as the Bharat Bachat application that seek to automate bookkeeping for self-employed individuals in emerging markets. A hybrid four-stage pipeline combining on-device OCR with domain-specific natural language correction is proposed for practical mobile deployment.