Multilingual Speech-to-Text Encoding System Based on Articulatory Spatial Coordinates and Phonation Temporal Dynamics: Reconstructing the Original Graphemic Principles of Hunminjeongeum for Next-Generation AI Tokenizers
Abstract Current Large Language Models (LLMs) and speech processing systems rely heavily on fixed subword tokenizers (e.g., BPE) or purely statistical byte-level patching (e.g., Byte Latent Transformer). While subword tokenizers suffer from severe token fragmentation and computational inequality across non-English languages, byte-level models lack inherent phonetic or articulatory information, requiring auxiliary entropy models and high memory overhead. This white paper introduces a novel, physically grounded AI encoding architecture derived from the reconstructed graphemic principles of Hunminjeongeum (1446). By rectifying historical misinterpretations of the Hunminjeongeum Haeryebon—specifically restoring the ancient seal/bronze script (Gu-zhuan, 字倣古篆) origins of basic graphemes (牙, 齒, 天, 立, 大), establishing a Dual-Anchor model for velar/molar sounds (ㆁ as temporal onset vs. ㄱ as graphemic base), and redefining stroke addition (Ga-hoek, 加劃) and gemination (Byeong-seo, 竝書) as temporal dynamics of articulatory speed (厲, rapidity/urgency) and duration (凝, prolongation)—this system encodes both spatial articulatory coordinates and temporal acoustic parameters directly into minimal graphemic tokens. This architecture significantly reduces vocabulary size, eliminates auxiliary boundary predictors, and enhances cross-lingual speech-text alignment.
Authors
- Dae-jong (朴大鍾) Bark
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24275245
- Primary Topic
- Speech Recognition and Synthesis
- Type
- preprint