StenoAssist: Smart Legal Sentence Suggestion System using LLMs

StenoAssist demonstrates a complete, functional end-to-end pipeline for automated shorthand recognition and legal document transcription, addressing a genuine and unmet need in India’s judicial system. The three-strategy recognition ensemble effectively handles the severe one-shot learning challenge inherent in the shorthand dataset. The two-pass LLM post-processing pipeline compensates for residual recognition errors through domain-specific contextual correction, achieving high-quality output even when the raw OCR confidence is moderate. The full-stack web application delivers a professional, deployable tool accessible to legal professionals without technical expertise. Future work will explore: (1) Vision Transformer (ViT) backbone replacing EfficientNet-B0 for improved symbol accuracy on ambiguous low-stroke symbols; (2) a React Native mobile application enabling on-the-go shorthand scanning via smartphone camera; (3) attention-based sequence-to-sequence decoding to eliminate the need for explicit per-symbol segmentation; (4) extension of the recognition system to Gregg and Teeline shorthand systems used internationally; and (5) cloud deployment with GPU inference acceleration and enterprise-grade authentication for court system integration.

[1] Tan, M. & Le, Q. V. (2019). EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. ICML 2019, PMLR 97, pp. 6105–6114. [2] Touvron, H. et al. (2023). Llama 2: Open Foundation and Fine-Tuned Chat Models. Meta AI Research. arXiv:2307.09288. [3] Nguyen, T. T. et al. (2021). Deep Learning for Visually Rich Document Understanding. ACM Computing Surveys, 54(6). [4] Pal, U., Wakabayashi, T., & Kimura, F. (2020). Handwritten Indian Script Recognition: A Survey. Springer PREMI Conference Proceedings. [5] Graves, A., Fernández, S., Gomez, F., & Schmidhuber, J. (2006). Connectionist Temporal Classification. ICML 2006, pp. 369–376. [6] Dosovitskiy, A. et al. (2020). An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale. ICLR 2021. [7] Wightman, R. (2019). PyTorch Image Models (timm). GitHub. https://github.com/huggingface/pytorch-image-models. [8] Groq Inc. (2024). Groq API Documentation. https://console.groq.com.