Improving Handwritten Stenography Recognition Using Domain Specific Encoding and Synthetic Data Augmentation

Handwritten form of stenography recognition is a challenging task due to high variations in shorthand symbols and a lack of annotated datasets. A strong baseline for stenographic text recognition, leveraging the newly created LION dataset of shorthand manuscripts. To alleviate symbol ambiguity as well as low-resource constraints in this domain, incorporating stenographic domain expertise is proposed through encoding schemes and synthetic data generation. In the experiments, a Gated-CNN-BGRU architecture is used, trained with Connectionist Temporal Classification (CTC) loss. Pre-train on synthetic data and fine-tuning on real data greatly improve CER and WER. The experiments showed the effectiveness of stenography-specific encoding schemes in improving WER (Word Error Rate) over baseline models by up to 10% when using Melin-based encodings. This work opens pathways for transcribing historical stenographic manuscripts more effectively.