Every figure on this page is computed from experiments/logs/speed_vs_tv_all.csv
(600 trials, data dictionary). Run
python analysis/verify_results.py to recompute them and fail on any mismatch.
Design. Within-subject, 20 participants. Two conditions: KIH (glove) and a TV-style cursor on-screen keyboard (arrow-key navigation + select, the method used on smart TVs and in XR). Condition order counterbalanced across participants. Each condition: 10 minutes of training, then 15 one-minute trials. Words were drawn at random from a fixed 300-word pool, and participants typed without looking at the device.
Metrics. Speed as WPM = jamo-per-minute ÷ 5 (so figures are comparable with the international literature). Accuracy: MSD-based uncorrected error rate for KIH, and excess-keystroke (correction) rate for the cursor keyboard — the two definitions are not interchangeable, so accuracy is compared as a trend, not head-to-head. A corrected KSPC (excluding backspace and submit keys) was used to check that the designed keystroke cost reproduces in real use.
| KIH (glove) | TV on-screen keyboard | |
|---|---|---|
| Speed, trial 1 | 2.3 WPM | 5.6 WPM |
| Speed, trial 15 | 10.6 WPM (4.6×) | 7.5 WPM |
| Power-law fit | WPM = 1.55·n0.602, R² = 0.636 | WPM = 5.80·n0.079, R² = 0.142 |
| Accuracy, trial 1 → 15 | 17.2% → 4.1% uncorrected error (−76%) | 4.3% → 5.6% correction rate |
| Crossover | passes the baseline at ≈ trial 9 | — |
Reading the fits. The exponent is the whole story: 0.602 against 0.079. The baseline was already saturated at first contact — participants arrive knowing how to drive a cursor keyboard, and that is as fast as the method gets (its R² of 0.142 says the "learning curve" is mostly noise). KIH starts less than half as fast and is still climbing at trial 15. Extrapolating the fit gives ≈12 WPM at 30 trials and ≈16 WPM at 50.
No speed–accuracy trade-off. Both improved together: 4.6× faster while uncorrected errors fell 76%. The usual concern about a new input method — that early speed is bought with errors — does not appear here.
Context. 10.6 WPM after 15 minutes of practice sits with recent hands-/eyes-free techniques (ankle gestures 11–13 WPM, CHI 2026; head gestures 9.8 WPM) and above the ~8 WPM reported for commercial cursor-selection keyboards — while leaving gaze and posture unconstrained, which none of those baselines do simultaneously.
Corpus: 199,806 keystroke-level jamo from modern Korean prose, cross-checked against the National
Institute of Korean Language frequency survey (mapping_analysis.py).
| Metric | Value | Note |
|---|---|---|
| Theoretical KSPC (taps per jamo) | 1.269 | multi-tap layout |
| KSPC of a frequency-blind uniform layout | 1.727 | baseline |
| Reduction | −26.5% | from the base-jamo = 1 tap structure |
| Input completed in a single tap | 74.2% | |
| Adjacent jamo pairs sharing a button | 3.60% | only these pay the multi-tap timeout |
| Most frequent same-button pair | ㅡ+ㅣ (syllable 의) | ≈32% of all same-button pairs |
The layout was designed for learnability first — follow Hangul's own 가획·병서 derivation and the only rule to memorize is "same family, same button, press again." The efficiency above is a by-product: Hangul's base letters largely coincide with its highest-frequency jamo, so a derivation-faithful layout is also close to a frequency-optimal one.
The multi-tap engine is verified on the host without hardware
(make -C firmware/keyboard_glove/test both, 16 checks × 2 build configurations):
- a tap sequence commits immediately once no further tap can extend it (3 taps for most consonants, 2 for vowels, 1 for function keys) — no timeout wait
- a sequence interrupted by a different button commits early rather than merging (ㄱ+ㅏ+ㄱ typed quickly stays ㄱㅏㄱ instead of collapsing to ㅋㅏ)
- switch chatter inside the debounce window is not counted as a tap
- ㅢ (ㅡ = 2 taps, then ㅣ = 1 tap on the same button) resolves correctly
The tap window itself is not a fixed constant: it is measured per user by
tap_calibration.py, which labels every inter-tap interval as intentional multi-tap or separate
keystroke and picks the threshold that minimises misclassification.