Survey Report on StenoAssist: Smart Legal Sentence Suggestion System using LLMs

In the legal domain, stenographers and court reporters capture spoken proceedings using shorthand notes. These notes often contain symbols, abbreviations and non-standard orthography, making them difficult for non-experts to read. Automating the conversion of shorthand notes into legally intelligible sentences could save time, reduce transcription errors and enable more efficient access to court records. This survey analyses research on three key components of such a system: optical character recognition (OCR) for handwritten shorthand, methods for expanding abbreviations into full legal phrases and the use of large language models (LLMs) to improve sentence quality. It reviews convolutional recurrent neural network (CRNN) architectures for scene and handwritten text recognition, discusses the connectionist temporal classification (CTC) loss used for unsegmented sequence labelling and summarises current approaches to abbreviation expansion and legal language modelling. The survey situates these technologies within StenoAssist—a hybrid system that recognizes shorthand images, decodes them with a lexicon and expands them into legal sentences using a glossary and an LLM. We evaluate the CRNN–lexicon pipeline on the Kaggle shorthand dataset and discuss the benefits and limitations of integrating LLMs for legal text.