3-Key-Input paper finds 3 keys + GPT-4o viable for text entry

A new paper from Naoki Kimura, presented at ICASSP 2026, explores the theoretical minimum number of physical keys needed for effective text entry when paired with modern language models. The study, titled "3-Key-Input: Exploring the Theoretical Minimum Keys for Text Entry," systematically evaluates systems using 2 to 5 keys combined with different letter-to-key mappings (layout-based, frequency-based, and intentionally worst-case) and decoders (Trie-only, GPT-2 beam search, GPT-4o selection). On a 300-sentence English corpus spanning business, conversational, and technical domains, the results are striking: three keys plus GPT-4o achieves a character error rate (CER) of 9.46% and a word error rate (WER) of 12.20% – a 59% relative reduction in CER compared to two keys (23.3% CER).

While increasing to five keys further improves accuracy (CER 5.4%), the marginal gains diminish, suggesting three keys represent a practical sweet spot. Mapping choice has surprisingly little impact (ΔCER < 0.5 pp), and even intentionally poor mappings degrade performance only slightly. However, technical sentences see roughly twice the error rate of business texts, highlighting domain dependence. The key-stream entropy at three keys is 1.54 bits/char. These findings open up new possibilities for assistive devices and ultra‑constrained mobile form factors, where reducing physical keys greatly increases design freedom without sacrificing usable input speed.