Quantifying Perceived Accessibility in Non-visual Interaction
(CHI'23)
Problem:
Blind screen reader users rely entirely on keyboard navigation to interact with desktop applications. However, navigating through menus, toolbars, and dialogs is often tedious and inconsistent across applications. Traditional accessibility evaluation methods either rely on developer compliance checks or require time-consuming user studies, and none provide a scalable way to estimate 'perceived accessibility'—how accessible a software feels to blind users—based on real-life usage. This gap makes comparing or improving applications from a non-visual interaction perspective difficult.
Solution:
We developed a probabilistic framework to estimate the perceived accessibility of desktop software using keystroke-based screen reader navigation. The framework was inspired by the findings of a formative study with 11 screen reader users. The formative study findings are available in the figures on the left.
Our work in this project:
Builds a probabilistic interaction model from UI hierarchies extracted via accessibility APIs, without needing pixel data or screen recordings.
Defines three task-agnostic, user-aligned metrics:
Complexity – expected keystrokes to reach a target (see Fig. 10).
Coverage – percentage of interface reachable within k keystrokes.
Reachability – fraction of the interface that can be accessed efficiently with shortcuts.
Validates these metrics through another study with the same 11 participants as the formative study.
Benchmarks 11 widely used Windows applications, including Notepad, Spotify, and Microsoft Word (see Fig. 10).
Our metrics revealed clear differences in perceived accessibility between applications:
Notepad had a low complexity score (1.63), while MS Word had a significantly higher score (3.51), aligning with participant preferences (see Fig. 10).
Complexity, Coverage, and reachability metrics highlighted how UI structure design dramatically affects interface accessibility (see the figures on the left).
Participants described the metrics as intuitive, reflective of real effort, and consistent with their experiences.
Applications with high complexity and low coverage were described as frustrating and challenging, confirming metric validity.
This work offers the first automated, model-based method for estimating perceived accessibility of desktop apps from a blind user's perspective. Its contributions include:
Enabling developers and auditors to quantify and compare accessibility without human testing (see the figures on the left).
Highlighting concrete areas for improvement, such as reducing navigation complexity or adding effective shortcuts.
Providing blind users with data-driven tools to choose more usable applications.
Shifting accessibility evaluation toward scalable, user-centered, and interface-aware methods.