Azimuth: Designing Accessible Dashboards for Screen Reader Users
Srinivasan, Arjun, Harshbarger, Tim, Hilliker, Darrell, Mankoff, Jennifer · 2023 · Proceedings of the 25th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS) · doi:10.1145/3597638.3608405
Summary
This paper presents Azimuth, a prototype system that converts JSON-based dashboard specifications into web-based dashboards optimized for screen reader navigation, accompanied by automatically generated textual descriptions to support comprehension and interaction. Dashboards — visual interfaces combining charts, key performance indicators (KPIs), filtering widgets, and interactive data querying — are ubiquitous in business, healthcare, public policy, and other domains, yet they are largely inaccessible to blind and low vision (BLV) users. Prior research has addressed making individual charts accessible through alt text, sonification, and structured navigation, but the multi-component, interactive nature of dashboards introduces distinct challenges: unpredictable component ordering, hidden interactivity between charts, implicit filtering through click actions, and the difficulty of tracking data changes across multiple linked visualizations. The research followed a participatory co-design process with two blind screen reader users (who became co-authors) over four months of weekly sessions, iterating on dashboard prototypes modeled on real-world examples from Tableau Public and Microsoft Power BI. This process yielded six design goals addressing dashboard structure (consistent DOM hierarchy with heading-based navigation), explicit filter presentation (surfacing implicit chart interactions as visible filter widgets), textual descriptions (summarizing structure, interactivity, and key data facts), screen-reader-optimized description formatting (using headings, bullet lists, and hyperlinks), change summaries after filter interactions, and support for multiple analytic modes (subset analysis vs. comparison). Azimuth is implemented in JavaScript using HighCharts.js for rendering and includes a reusable API that can also augment existing Tableau or Vega-Lite dashboards with accessible descriptions.
Key findings
A user study with five BLV participants (four using JAWS, one using VoiceOver) evaluated the generated dashboards across targeted analysis tasks and open-ended exploration. Participants successfully used the dashboards to answer specific data questions — three of five completed all eight targeted tasks, with correct response rates ranging from 100% (P1, P5) to 86% (P2). The dashboard descriptions were unanimously valued: participants read them during both training and task phases, using the layout summary to build mental models of the dashboard structure and the data facts to identify key takeaways. Participants employed diverse navigation strategies — heading-based browsing, search-based jumping, and VoiceOver item chooser — demonstrating that the consistent heading structure supported flexible screen reader interaction rather than enforcing a single navigation path. The change description feature was particularly impactful for interactive querying: participants used it to sift through data subsets by applying filters and reading the updated summaries, with three of five preferring the comparison mode that contextualized changes relative to the previous dashboard state. During open-ended exploration, all participants independently formulated and investigated their own questions, a capability they described as novel. Participant feedback was overwhelmingly positive (average 4.6/5 across six evaluation dimensions), with one participant noting: "Compared to having to download data and re-run updates this will save a lot of time but the main reason this is good though, is because I should not have to download the data and do all that. I should be able to have the same access, like everyone else."
Relevance
This research addresses a critical gap in data accessibility: while individual chart accessibility has received growing attention, dashboards — the primary way organizations share and monitor data — remain largely inaccessible to screen reader users, forcing BLV professionals to either work with raw data exports or rely on sighted colleagues for interpretation. Azimuth demonstrates that accessible dashboards require more than accessible charts; they need structured navigation hierarchies, explicit interactive affordances, and dynamically generated textual summaries that track data changes. For practitioners, the six design goals provide immediately applicable guidelines for dashboard development: use consistent heading structures, surface implicit interactions as explicit filter widgets, generate descriptions with data facts organized by statistical salience, and provide change summaries after filter actions. The participatory co-design methodology — where two blind individuals became full co-authors — also models best practice for inclusive research. The finding that the system promoted user autonomy and independent data exploration, rather than just answering predefined questions, underscores the importance of designing for open-ended analysis rather than merely providing static alternative text.
Tags: data visualization · screen readers · dashboards · blind and low vision · co-design · text generation · accessible navigation · data analysis
Standards referenced: WCAG