Learning to Recognize Handwriting Input with Acoustic Features
summary
A mobile app records pen-on-surface sound, uses adaptive segmentation to slice strokes, extracts STDFT features, and sends fragments to a server running an Inception-LSTM network to classify letters. Prototype uses app–cloud architecture, reports segmentation rates of 99.76% (letters) and 99.16% (words), per-letter accuracy 73.8%, word top-1 88.5% and top-5 94.17%, and claims 93.2% accuracy for new users in practical scenarios.
AI-written summary · about