Typing Performance of Blind Users: An Analysis of Touch Behaviors, Learning Effect, and In-Situ Usage
Hugo Nicolau, Kyle Montague, Tiago Guerreiro, André Rodrigues, Vicki L. Hanson · 2015 · ASSETS '15: Proceedings of the 17th International ACM SIGACCESS Conference on Computers & Accessibility · doi:10.1145/2700648.2809861
Summary
This longitudinal study provides the most detailed analysis to date of how novice blind users learn to type on touchscreen keyboards using the Explore by Touch paradigm (where users slide fingers across the screen to hear key names, then lift to select). Five legally blind participants who had never owned a smartphone used Samsung S3 devices with TalkBack for eight weeks, with researchers collecting both in-situ usage data (background logging) and conducting weekly laboratory typing assessments. The study extends typical text-entry research by analyzing not just aggregate speed and accuracy metrics, but character-level error types and detailed touch behaviors including initial contact points, exploration movements, and lift positions. Over eight weeks, participants entered 32,764 characters during 51 hours of active typing. The unconstrained text-entry protocol allowed participants to correct errors as they would naturally, enabling analysis of both uncorrected and corrected error rates. The researchers were motivated by a critical gap: while prior work established that non-visual touchscreen input is slow and error-prone compared to sighted typing, there was little understanding of why errors occur or how to improve performance. By examining the full trajectory of each keystroke—from initial touchdown through exploration to final selection—the study reveals the mechanics underlying both successful input and errors.
Key findings
Users improved significantly but slowly, from 1.6 WPM in week one to 4.0 WPM by week eight (0.3 WPM per week). Total error rates dropped from 26% to 7.4%. Extrapolating the learning curve suggests users would reach 5 WPM by week 16. Participants spent 32% of their time correcting errors initially, dropping to 13% by week eight. Substitutions (typing the wrong character) are the dominant error type, consistently higher than insertions or omissions. Unlike sighted typists who show predictable directional offset patterns, blind users' touch points scatter over intended keys and near their edges with no clear directional bias. This has implications for adaptive keyboard algorithms—existing touch models trained on sighted users may not transfer. The most actionable finding concerns speech feedback timing: 64% of substitution errors occurred when users' fingers crossed the intended key during exploration but failed to lift in time—the mismatch between touch position and audio feedback caused them to select the wrong key. By week eight, only 36% of substitutions were true "slip errors" (lifting on adjacent keys); the rest resulted from various causes including accidental touches, phonetically similar keys (M-N confusions persisted at 4.5%), and overconfidence in spatial models. Keys near physical device edges (Q, A, P, L) achieved 71-75% accuracy, while center keys (B, M) managed only 14-16%. The space bar reached 99% accuracy due to its size and bottom-edge positioning. Users developed an "emergent keyboard" mental model shifted toward the bottom and center of the screen, with significant horizontal overlap between adjacent keys (especially M and N).
Relevance
This research provides essential empirical grounding for designing better non-visual touchscreen keyboards. The finding that speech feedback delay causes the majority of substitution errors directly suggests that synchronizing audio output with touch position should be a primary design goal—when users hear "M" they should be touching M, not having just left it. Several design implications emerge: (1) Filter accidental touches by detecting characteristically short movement times and distances; (2) Leverage movement trajectory data to predict intended targets rather than relying solely on lift position; (3) Implement language-based correction for omission errors, since 68% go undetected and uncorrected; (4) Adaptive keyboards must account for the lack of consistent directional bias in blind users' touches. The detailed touch behavior data also reveals that correction is time-consuming and inefficient—users delete correct characters one-third of the time while attempting to fix errors, and they never used cursor movement operations. This suggests auto-correct and intelligent suggestions could dramatically improve efficiency, though they must be carefully designed for non-visual use. The longitudinal perspective demonstrates that while learning does occur, it is slow enough that design improvements remain critical even for experienced users.
Tags: visual impairment · touchscreen · text entry · mobile accessibility · screen reader · keyboard · longitudinal study · TalkBack