How Chrome Accessibility Works

This document explains the technical details behind Chrome accessibility code by starting at a high level and progressively adding more levels of detail.

Please read the accessibility overview first.

Accessibility for a simple (non-browser) application

As described in the overview, every platform has its own accessibility APIs that are used by both assistive technology and sometimes by automation software. To better understand the challenges of accessibility support in Chromium, let‘s first explore what it’s like to build an accessible application using a standard UI toolkit.

Examples of standard toolkits would be Win32 UI controls or .NET components on Windows, or Cocoa NSViews on macOS, or Android Views via Java or Kotlin. When you use such a toolkit to build an application, a lot of accessibility comes for free.

Typically every UI element - including containers and grouping elements - gets its own corresponding accessibility element, implementing that platform‘s accessibility API. So if you add a button, checkbox, scroll container, or text field to your application, you don’t need to do any work to make it accessible - the built-in UI toolkit has provided all of that already.

Simple accessibility tree

Changing accessibility properties

There are some cases where a bit of extra information may be required on the part of the app author. If you add an image button, you may need to ensure you provide an accessible label (“alt text”). Other common simple modifications might include hiding UI elements from accessibility because they‘re primarily decorative, or marking a UI element as an alert to ensure it’s announced to screen reader users when it appears. However, none of these require writing more than a few lines of code.

Windows

The dynamic annotation API lets you change any accessibility property of a UI element represented by an HWND. Here's a tiny code snippet showing how you could override the accessible name property (PROPID_ACC_NAME) to be the text “Accessible Name”:

CComPtr<IAccPropServices> pAccPropSrv;
pAccPropSrv.CoCreateInstance(CLSID_AccPropServices);
COleVariant varName("Accessible Name");
pAccPropSrv->SetHwndProp(hwnd, OBJID_CLIENT, 0, PROPID_ACC_NAME, varName);

Android

On Android, if you don't want to subclass a View you can just set its accessibility delegate as an easy way to override one or more accessibility attributes. In this case we override the content description to say “Accessible Name”.

view.setAccessibilityDelegate(new AccessibilityDelegate() {
    public void onInitializeAccessibilityNodeInfo(View host, AccessibilityNodeInfo info) {
        super.onInitializeAccessibilityNodeInfo(host, info);
        info.setContentDescription("Accessible Name");
    }
});

Custom controls and subclassing

When building a custom control, more work is often required. As an example, suppose an application is implementing a “cover flow” image picker, such as the one found in the original iTunes album picker. Conceptually the user is picking one album from a large list, but visually it's shown as a horizontally scrolling list of album cover images that fly by in three dimensions.

Screenshot of a cover flow image picker

While this control may be implemented as a single UI element, from the perspective of accessibility APIs it would typically be represented as a parent element with a “list box” role and children with “option” roles. This can't be done using the code snippets above, which are only sufficient for overriding some simple accessibility properties. A more complex custom control like this requires subclassing.

To make the CoverFlow accessible, we would need to first need to subclass the appropriate accessibility interface, such as IAccessible on Windows or AccessibilityDelegate on Android. In addition to setting properties such as the name and role (like ComboBox on Windows, or Choice on Android), we'd want to override any methods that are used to walk the tree, in order to give this object accessible children - one for each option the user can pick.

Figure showing a single CoverFlow control in the UI object tree, and a tree ofcontrols in the accessibility tree

Each option would be another instance of our subclass of an accessibility interface. Each one might need to implement a role, name, and very importantly, a bounding box representing the position of that option within the view's bounds.

A single-process browser for basic HTML only

Let‘s next discuss what’s needed in order to make a single-process web browser accessible. At the time Chrome was first released, this is more or less how other browsers such as Safari and Firefox worked. However, let‘s start with the simplifying assumption that we’re dealing with basic HTML only, and no CSS or ARIA.

From the perspective of accessibility APIs, the web contents is one big custom control. Modern web browsers do not build their UI using platform UI elements.

Years ago, some browsers may have rendered web pages using platform UI elements, but this doesn't scale well as most UI frameworks get bogged down after thousands of UI elements, whereas many web pages have hundreds of thousands of elements.

In addition, note that web form controls support CSS effects including unusual ones like 3-D transformations, whereas most platform UI controls don't support, e.g. a 45-degree rotated text box.

The web contents can be represented as a tree of elements. For the moment let‘s ignore details between the DOM and the final rendered layout, and let’s just assume that every DOM node corresponds perfectly to one node in the accessibility tree.

Essentially the web browser needs to create one accessible object for every DOM node. The accessible object subclasses that platform's accessibility interface, and its accessible role, name, and other properties can all be computed by querying the corresponding DOM node. When DOM nodes are created or destroyed, the corresponding accessible objects need to be updated accordingly.

Here‘s an example of a DOM tree and a corresponding accessibility tree. There’s a 1:1 correspondence in this example.

Figure showing a tree of DOM nodes for a web page and a correspondingtree of accessibility nodes

Objects in the accessibility tree are sometimes called “wrappers” in this design. Each accessible object “wraps” its corresponding DOM element. The DOM element contains all of the state, and the accessible object just implements the accessibility API based on the DOM element. This design avoids the need for DOM code to need to be tightly bound to accessibility code.

For efficiency, these accessible objects are sometimes built lazily. Initially there could just be one accessible object for the root of the web contents. If that object is ever queried and asked for its children, the accessible objects corresponding to the children of the root DOM element could be created on-demand, and so on. This is useful because many users don‘t have any features enabled that use accessibility APIs, so very little work would be done unless they’re utilized.

The overall system so far is pretty simple - the DOM tree is used to build the accessibility tree, which is used to communicate with the platform accessibility APIs.

System diagram showing a browser consisting of a DOM tree that's usedto build an accessibility tree, which is used to communicate withassistive technology using platform accessibility APIs.

A single-process browser with CSS and ARIA

In the previous example we assumed that there was a 1:1 correspondence between a DOM node and an accessible node. Unfortunately that doesn‘t hold true. Here are some examples of where that assumption doesn’t work.

  • CSS can have generated content, like list item markers (bullets or numbers) or text inserted before or after an element with pseudoselectors like ::before and ::after.
  • Some DOM subtrees are never displayed (like <head>) or are hidden using display:none
  • Some DOM nodes are hidden using visibility:hidden, but it's possible for some of their descendant nodes to be visible
  • aria-hidden can hide a subtree from the accessibility tree even if it's displayed visually
  • ARIA role=none or role=presentation hides a node but not its subtree
  • aria-owns can be used to reparent a node within the accessibility tree

So in practice, the accessibility tree is similar to the DOM tree, but with important differences. Objects in the accessibility tree usually wrap DOM nodes, but occasionally they wrap pseudoelements (like list markers or CSS generated content) or other layout objects with no corresponding node (such as images that are children of pseudoelements).

Here's an example where the DOM tree has list items and the corresponding accessibility tree has nodes for the list item markers:

Figure showing a DOM tree with an ordered list of two items and theaccessibility tree containing extra nodes for the generated list items

Here's a different example where some of the content is excluded from the accessibility tree using aria-hidden:

Figure showing a DOM tree with a sidebar with aria-hidden=true on itand the corresponding accessibility tree where the sidebar isn't presentat all

This new system diagram reflects the complexity of the system when you consider both HTML, CSS, and ARIA all affecting the accessibility tree.

System diagram showing a browser consisting of a DOM tree that alongwith CSS is used to build a Layout tree, then together the DOM and Layouttrees build an accessibility tree, which is used to communicate withassistive technology using platform accessibility APIs.

A multi-process browser

In the next section we'll explore how the system diagram needs to change in order to support a multi-process browser.

See How Chrome Accessibility Works, Part 2