Handwriting Synthesis for Low-Delay Predicted Character Display

The patent segments the character generation process into two independent parallel paths: (1) token-based synthesis using collected handwriting strokes for common characters, and (2) model-based synthesis using conversion models for less frequent characters. This segmentation allows simultaneous processing of different character types, eliminating sequential delays while maintaining input prediction functionality.

The patent performs preliminary action by pre-collecting and storing handwriting strokes for common characters in a stroke collection during the user's handwriting process. This pre-prepared stroke data can be quickly retrieved and synthesized without waiting for complete character-by-character processing, thereby reducing display delay while preserving prediction accuracy.

The patent applies local quality by differentiating the synthesis approach based on character frequency and availability of collected strokes. For common characters with available stroke collections, token-based synthesis is used for high speed. For less frequent characters without sufficient stroke data, model-based synthesis is employed to maintain accuracy. This localized adaptation of synthesis methods optimizes both speed and accuracy for different character types.

The patent creates a composite synthesis system that combines two different synthesis methods (token-based and model-based) into a unified framework. The system dynamically selects and combines appropriate methods based on the specific character being synthesized, leveraging the strengths of both approaches to achieve both high speed and high accuracy across all character types.