Chrome Accessibility on Android

This document covers some of the technical details of how Chrome implements its accessibility support on Android.

Chrome plays an important role on Android - not only is it the default browser, but Chrome powers WebView, which is used by many built-in and third-party apps to display all sorts of content. Android includes a lightweight implementation called Chrome Custom Tabs, which is also powered by Chrome. All of these implementations must be accessible, and the Chrome & Chrome OS Accessibility team provides the support to make these accessibility through the Android API.

Accessibility on Android is heavily used. There are many apps that hijack the Android accessibility API to act on the user's behalf (e.g. password managers, screen clickers, anti-virus software, etc). Because of this, roughly 16% of all Android users are running the accessibility code (even if they do not realize it).

As background reading, you should be familiar with Android Accessibility and in particular AccessibilityNodeInfo objects, AccessibilityEvent and AccessibilityNodeProvider.

WebContentsAccessibility

The main Java class that implements the accessibility protocol in Chrome is WebContentsAccessibilityImpl.java. This class acts as the AccessibilityNodeProvider (see above) for a given tab, and will provide the virtual tree hierarchy, preform actions on the user's behalf, and send events to downstream services for changes in the web contents.

This class differs in a few key ways from other platforms. First, it represents and entire web page, including all frames. The ids in the java code are unique IDs, not frame-local IDs. They are typically referred to as virtualViewId in the code and Android framework/documentation. Another key difference is the construction of the native objects for nodes. On most platforms, we create a native object for every AXNode in a web page, and we implement a bunch of methods on that object that assistive technology can query. Android is different - it‘s more lightweight in one way, in that we only create a native AccessibilityNodeInfo object when specifically requested, when an Android accessibility service is exploring the virtual tree. In another sense it’s more heavyweight, though, because every time a virtual view is requested we have to populate it with every possible accessibility attribute, and there are quite a few.

WebContentsAccessibilityImpl is “lazy” and “on demand”

Every Tab in the Chrome Android browser will have its own instance of WebContentsAccessibilityImpl. The WebContentsAccessibilityImpl object is created using a static Factory method with a parameter of the WebContents object for that tab. See constructor. (Note: There are a few exceptions to this pattern, for example when constructing the object without access to a WebContents instance, such as in the case of PaintPreview.)

Although the WebContentsAccessibilityImpl object has been constructed (and technically instantiated), it will not do anything until an accessibility service is active and queries the system. The base class of Java widgets View.java, has a method getAccessibilityNodeProvider. Custom views, such as the ContentView.java class in Chrome (which holds all the web contents for a single Tab), can override this method to return an instance of AccessibilityNodeProvider. If an app returns its own instance of AccessibilityNodeProvider, then AT will leverage this instance when querying the view hierarchy. The WebContentsAccessibilityImpl acts as Chrome's custom instance of AccessibilityNodeProvider so that it can serve the virtual view hierarchy of the web to the native accessibility framework.

The first time that getAccessibilityNodeProvider is called by the Android system, the WebContentsAccessibilityImpl will be initialized. This is why we consider it “lazy” and “on demand”, because although it has technically been constructed and instantiated, it does not perform any actions until AT triggered its initialization. See WebContentsAccessibilityImpl#getAccessibilityNodeProvider and the associated onNativeInit methods. The getAccessibilityNodeProvider method will only be called when an accessibility service is enabled, and so by lazily constructing only after this call, we ensure that the accessibility code is not being leveraged for users without any services enabled.

Once initialized, the WebContentAccessibilityImpl plays a part in handling nearly all accessibility related code on Android. This object will be where AT sends actions performed by users, it constructs and serves the virtual view hierarchy, and it dispatches events to AT for changes in the web contents. The WebContentsAccessibilityImpl object has the same lifecycle as the Tab for which it was created, and although it won't fire events or serve anything to downstream AT if the tab is backgrounded/hidden, the object will continue to exist and will not be destroyed until the Tab is destroyed/closed.

AccessibilityNodeInfo

The AccessibilityNodeInfo object is at the core of Android accessibility. This is a rather heavy object which holds every attribute for a given node (a virtual view element) as defined by the Android API. (Note: The Android accessibiltiy API has different attributes/standards than the web or other platforms, so there are many special cases and considerations, more on that below).

As an AccessibilityNodeProvider, the WebContentsAccessibilityImpl class must override/implement the createAccessibilityNodeInfo method. This is the method that the accessibility framework will query on behalf of AT to understand the current virtual view hierarchy. On other platforms, the native-side code may contain the entire structure of the web contents in native objects, but on Android the objects are created “on demand” as requested by the framework, and so they are typically generated synchronously on-the-fly.

The information to populate the AccessibilityNodeInfo objects is contained in the accessibility tree in the C++ code in the shared BrowserAccessibilityManager. For Android there is the usual BrowserAccessibilityManagerAndroid, and the BrowserAccessibilityAndroid classes, as expected, but there is also an additional web_contents_accessibility_android.cc class. This class is what allows us to connect the Java-side WebContentsAccessibilityImpl with the C++ side manager, through the Java Native Interface (JNI).

When WebContentsAccessibilityImpl#createAccessibilityNodeInfo is called for a given virtual view (web node), the WebContentsAccessibilityImpl object calls into the native C++ code through JNI, connecting to web_contents_accessibility_android.cc. The web_contents_accessibility_android object in turn compiles information about the requested node from BrowserAccessibilityAndroid and BrowserAccessibilityManagerAndroid and then calls back into WebContentsAccessibilityImpl, again through the JNI, to send this information back to the Java-side to be populated into the AccessibilityNodeInfo object that is being constructed.

These roundtrips across the JNI come with an inherent cost. It is minuscule, but for thousands of nodes on a page, each with 25+ attributes, it would be too costly to make so many trips. However, passing all the attributes in one giant function call is also not ideal. We try to strike a balance by grouping like attributes together (e.g. all boolean attributes) into a single JNI trip, and make just a few JNI trips per AccessibilityNodeInfo object. These trips can be found in the WebContentsAccessibilityAndroid::PopulateAccessibilityNodeInfo method, which is called from WebContentsAccessibilityImpl#createAccessibilityNodeInfo, and is the core method for compiling a node's attributes and passing them to the Java-side.

Java-side caching mechanism

One of the most significant performance optimizations in the Android accessibility code is the addition of a caching mechanism for the Java-side AccessibilityNodeInfo objects. The cache is built as a simple SparseArray of AccessibilityNodeInfo objects. We use a SparseArray instead of a HashMap because on Java the HashMap requires Objects for both the key and value, and ideally we would use the virtualViewId of any given node as the key, and this ID is an int (primitive type) in Java. So the SparseArray is more light weight and is as efficient as using the HashMap in this case. The array contains AccessibilityNodeInfo objects at the index of the node's corresponding virtualViewId. If an invalid ID is requested, null is returned.

In WebContentsAccessibilityImpl's implementation of createAccessibilityNodeInfo, the cache is queried first, and if it contains a cached version of an object for this node, then we update that reference and return it. Otherwise the object is created anew and added to the cache before returning. The rough outline of the code is:

private SparseArray<AccessibilityNodeInfoCompat> mNodeInfoCache = new SparseArray<>();

@Override
public AccessibilityNodeInfoCompat createAccessibilityNodeInfo(int virtualViewId) {
  if (mNodeInfoCache.get(virtualViewId) != null) {
    cachedNode = mNodeInfoCache.get(virtualViewId);
    // ... update cached node through JNI call ...
    return cachedNode;
  } else {
    // Create new node from scratch
    AccessibilityNodeInfo freshNode = // ... construct node through JNI call.
    mNodeInfoCache.put(virtualViewId, freshNode);
    return freshNode;
  }
}

When returning a cached node, there are some fields that we always update. This requires a call through the JNI, but it still much more efficient than constructing a full node from scratch. Rather than calling PopulateAccessibilityNodeInfo, we call WebContentsAccessibilityAndroid::UpdateCachedAccessibilityNodeInfo. This method updates the bounding box for the node so that AT knows where to draw outlines if needed. (Note: it also technically updates RangeInfo on some nodes to get around a bug in the Android framework, more on that below.)

We clear nodes from the cache in BrowserAccessibilityManager::OnNodeWillBeDeleted. We also clear the parent node of any deleted node so that the AccessibilityNodeInfo object will receive an updated list of children. We also clear any node that has a focus change during FireFocusEvent.

Bundle extras / API gaps

Much of the richness of the web cannot be captured by the Android accessibility API, which is designed from a native (Java) widget perspective. When there is a piece of information that an AT would like to have access to, but there is no way to include it through the standard API, we put that info in the AccessibilityNodeInfo's Bundle.

The Bundle, accessed through getExtras() is a map of key-value pairs which can hold any arbitrary data we would like. Some examples of extra information for a node:

  • Chrome role
  • roleDescription
  • targetUrl

We also include information unique to Android, such as a “clickableScore”, which is a rating of how likely an object is to be clickable. The boolean “offscreen” is used to denote if an element is “visible” to the user, but off screen (see more below). We include unclipped bounds to give the true bounding boxes of a node if we were not clipping them to be only onscreen. The full list can be seen in the list of constants at the top of WebContentsAccessibilityImpl.java.

Asynchronously adding “heavy” data

Sometimes apps and downstream services will request we add additional information to the AccessibilityNodeInfo objects that is too computationally heavy to compute and include for every node. For these cases, the Android API has a method that can be called by AT, addExtraDataToAccessibilityNodeInfo. The method is part of the AccessibilityNodeProvider, and so WebContentsAccessibilityImpl has its own implementation of this for Chrome. When called with valid arguments, this will start an asynchronous process to add this extra data to the given AccessibilityNodeInfo object. The two current implementations of this are to add text character locations, and raw image data.

AccessibilityNodeInfoCompat

The Android framework includes a support library for backwards compatibility. This is a mechanism that allows Android to add new attributes or methods to their API while also including backwards compatibility across previously released versions of Android. The WebContentsAccessibilityImpl class makes heavy use of this to ensure all features are supported on older versions of Android. This was a recent change, and the rest of the Chrome code base still uses the non-Compat version of the accessibility code because there is no use-case yet to switch. To make the change as minimal as possible, the WebContentsAccessibilityImpl uses the Compat version internally, but when communicating with other parts of Chrome, it will unwrap the non-Compat object instead. For this entire document, whenever AccessibilityNodeInfo is mentioned, technically speaking we are using an AccessibilityNodeInfoCompat object, and we use the AccessibilityNodeProviderCompat version as well. This library is designed to be transparent to the end-user though, and for simplicity we generally do not include the word ‘Compat’ in documentation, conversation, etc.

Responding to user actions

As the AccessibilityNodeProvider, the WebContentsAccessibilityImpl is responsible for responding to user actions that come from downstream AT. It is also responsible for telling downstream AT of Events coming from the web contents, to ensure that AT is aware of any changes to the web contents state.

performAction

One of the most important methods in WebContentsAccessibilityImpl is the implementation of performAction. This method is called by the Android framework on behalf of downstream AT for any user actions, such as focus changes, clicks, scrolls, etc. For most actions, we call through the JNI to web_contents_accessibility_android, which will call a corresponding method in BrowserAccessibilityManager to send this request to the underlying accessibility code. The performAction method parameters include a virtualViewId, the action, and a Bundle of args/extras. The Bundle may be null, but sometimes carries necessary information, such as text that a user is trying to paste into a field. If the accessibility code is able to successfully act on this performAction call, it will return true to the framework, otherwise false.

AccessibilityEventDispatcher

The other direction of communication is from the accessibility code into the framework and downstream AT. For this we dispatch an AccessibilityEvent. Often times a call to performAction is paired with one or more AccessibilityEvents being dispatched in response, but AccessibilityEvents can also be sent without any user interaction, but instead from updates in the web contents. The AccessibilityEvents are relatively lightweight, and they are constructed following the same model as the AccessibilityNodeInfo objects (i.e. calling into web_contents_accessibility_android and being populated through a series of JNI calls). However, the events can put significant strain on downstream AT, and this is where another important performance optimization was added.

Traditionally an app would realize it needs to generate and send an AccessibilityEvent, it would generate it synchronously and send it on the main thread. The web is more complicated though, and at times could be generating so many events that downstream AT is strained. To alleviate this, we have implemented a throttling mechanism, the AccessibilityEventDispatcher. Whenever WebContentsAccessibilityImpl requests to send an AccessibilityEvent, it first goes through this dispatcher. For most events, they are immediately passed along to the Android framework (e.g. user clicks, focus changes). Some events are instead put in a queue and dispatched or dropped based on future user actions.

For events that we wish to throttle, we will immediately send the first event received of that type. We record the system time of this event. If another event request of the same type is received within a set time limit, which we call the throttleDelay, then we will wait until to send that event until after the delay.

For example, scroll events can only be sent at most once every 100ms, otherwise we would attempt to send them on every frame. Consider we have sent a scroll event and recorded the system time to start a throttle delay. If we try to dispatch another scroll event in that delay window, it will be queued to release after the delay. If during that throttleDelay period another scroll event is added to the queue, it will replace the previous event and only the last one added in the 100ms window is dispatched. The timer would then restart for another 100ms.

The delay times can be specific to a view, or specific to an event. That is, we could say “only dispatch scroll events every 100ms for any given view”, meaning two different nodes could send scroll events in close succession. Or we can say “only dispatch scroll events every 100ms, regardless of view”, in which case all views trying to send that event type will enter the same queue.

The event types and their delays can be found in the constructor of the WebContentsAccessibilityImpl, which also includes the construction of the AccessibilityEventDispatcher.

Delayed AccessibilityEvent construction

The final performance optimization related to events (that is released on 100% stable), was to delay the construction of the AccessibilityEvent until it is about to be dispatched. Implementing the Dispatcher helps a significant amount, but there are many events that can be dropped, and there is no reason to construct an event until we are sure it will be dispatched. So we do not construct the AccessibilityEvents until the moment the Dispatcher has started the request to send the event to the Android framework.

Testing on Android

Testing on Android happens through a couple of build targets depending on what it is that we want to test. Android has tests present in the content_browsertests target, same as the other platforms, which tests the BrowserAccessibilityManagerAndroid and BrowserAccessibilityAndroid through the various DumpAccessibilityTreeTests and DumpAccessibilityEventsTests. However, these tests do not cover the web_contents_accessibility_android layer, or any of the Java-side code. The web_contents_accessibility_android object and the associated WebContentsAccessibilityImpl object are not created for content_browsertests and require a full browser instance to be available (or at least the content shell). To handle these types of tests we must use the content_shell_test_apk target, which will run an instance of a web contents and allow the creation/execution of WebContentsAccessibilityImpl and the corresponding native object. And finally there is the chrome_public_test_apk, which is used to test the Chrome Android UI, outside the web contents, which is necessary for testing accessibility features that have a user-facing Android UI, such as image descriptions, the accessibility settings pages, and page zoom.

Testing the “missing layer”

The “missing layer” in testing refers to the gap in testing for WebContentsAccessibilityImpl, and namely web_contents_accessibility_android mentioned above. There are three main classes we use to test these. They are:

  • WebContentsAccessibilityTest

    This test suite is used to test the methods of WebContentsAccessibilityImpl.java. It tests the various actions of performAction, construction of AccessibilityEvents, and various helper methods we use throughout the code.

  • WebContentsAccessibilityTreeTest

    This class is the Java-side equivalent of the DumpAccessibilityTreeTests. This test suite opens a given html file (shared with the content_browsertests), generates an AccessibilityNodeInfo tree for the page, and then dumps this tree and compares with an expectation file (denoted with the ...-expected-android-external.txt suffix). We continue to keep around the content_browsertests because a failure in one and not the other would provide insight into a potential bug location.

  • WebContentsAccessibilityEventsTest

    This class is the Java-side equivalent of the DumpAccessibilityEventsTests. Same as the suite above, it shares the same html files as the content_browsertests, opens them, runs the Javascript, and records the AccessibilityEvents that are dispatched to downstream AT. There is no Android version of the DumpAccessibilityEventsTests though, so these expectation files are suffixed with the usual ...-expected-android.txt.

When new tests are added for content_browsertests, the associated test should also be added in WebContentsAccessibility*Test, and there are PRESUBMIT warnings to remind developers of this (although they are non-blocking).

Writing new tests

Adding tests is as easy on Android as it is on the other platforms because the mechanism is in place and only a single new method needs to be added for the test.

If you are adding a new events test, “example-test.html”, you would first create the html file as normal (content/test/data/accessibility/event/example-test.html), and add the test to the existing dump_accessibility_events_browsertests.cc:

IN_PROC_BROWSER_TEST_P(DumpAccessibilityEventsTest, AccessibilityEventsExampleTest) {
  RunEventTest(FILE_PATH_LITERAL("example-test.html"));
}

To include this test on Android, you would add a similar block to the WebContentsAccessibilityEventsTest.java class:

@Test
@SmallTest
public void test_exampleTest() {
    performTest("example-test.html", "example-test-expected-android.txt");
}

Some tests on Android won't produce any events. For these you do not need to create an empty file, but can instead make the test line:

    performTest("example-test.html", EMPTY_EXPECTATIONS_FILE);

The easiest approach is to use the above line, run the tests, and if it fails, the error message will give you the exact text to add to the -expected-android.txt file. The -expected-android.txt file should go in the same directory as the others (content/test/data/accessibility/event).

For adding a new WebContentsAccessibilityTreeTest, you follow the same method but include the function in the corresponding Java file.

Note: Writing WebContentsAccessibilityTests is much more involved and there is no general approach that can be encapsulated here, same with the UI tests. For those there are many existing examples to reference, or you can reach out to an Android developer.

Running tests and tracking down flakiness

Running tests for Android can seem a bit daunting because it requires new build targets, an emulator, and different command-line arguments and syntax. But Android has a few nifty tricks that don‘t exist on every platform. For example, Android tests can be run on repeat indefinitely, but set to break on their first failure. This is great for tracking down flakiness. It is also possible to use a local repository to test directly on the build bots, which is great when a test works locally but flakes or fails during builds. Let’s look at some basic examples.

First ensure that you have followed the basic Android setup guides and can successfully build the code. You should not proceed further until you can successfully run the command:

autoninja -C out/Debug chrome_apk

One of the most important things to remember when building for unit tests is to use the x86 architecture, because most emulators use this. (Note: For running on try bots however, you'll want arm64, more on that below). Your gn args should contain at least:

target_os = "android"
target_cpu = "x86"

To run the types of tests mentioned above, you'll build with a command similar to:

autoninja -C out/Debug content_shell_test_apk

The filtering argument for tests is -f rather than the --gtest_filter that it is used with content_browsertests. So to run an example WebContentsAccessibilityTreeTest test, you may use a command such as:

out/Debug/bin/run_content_shell_test_apk --num_retries=0 -f "*WebContentsAccessibilityTreeTest*testExample"

This would look for an x86 phone to deploy to, which should be your emulator. You can choose to setup an emulator in Android Studio, or you can use some of the emulators that come pre-built in the repo. More information here. In general it is best to run on one of the newer Android versions, but sometimes the newest is unstable. To specify an emulator to use, you include the --avd-config argument, along with the desired emulator (see link above for full list). This will run the test without opening a window, but if you'd like to see an emulator window you can add the --emulator-window argument. The --repeat=# argument allows repeats, and if set to -1 along with the --break-on-failure argument, the test will run repeatedly until it fails once.

Putting this all together, to run the example test with no retries per run, run repeatedly until failure, on the Android 11 emulator, with a window available, you would use the command:

out/Debug/bin/run_content_shell_test_apk \
    --num_retries=0 \
    --repeat=-1 \
    --break-on-failure \
    --emulator window \
    --avd-config tools/android/avd/proto/generic_android30.textpb \
    -f "*WebContentsAccessibilityTreeTest*testExample"

All of this information also applies to the UI tests, which use the target:

autoninja -C out/Debug chrome_public_test_apk

In this case we would have a similar command to run an ImageDescriptions or Settings related test:

out/Debug/bin/run_chrome_public_test_apk \
    --num_retries=0 \
    --repeat=-1 \
    --break-on-failure \
    --emulator window \
    --avd-config tools/android/avd/proto/generic_android30.textpb \
    -f "*ImageDescriptions*"

Running on try bots

For more information you should reference the mb.py user guide.

Note: When running on the trybots, you often need to use target_cpu = "arm64", since these are actual devices and not emulators.

It is not uncommon when working on the Android accessibility code to have a test that works fine locally, but consistently fails on the try bots. These can be difficult to debug, but if you run the test directly on the bots it is easier to gain insights. This is done using the mb.py script. You should first build the target exactly as outlined above (or mb.py will build it for you), and then you use the mb.py run command to start a test. You provide a series of arguments to specify which properties you want the try bot to have (e.g. which OS, architecture, etc), and you also can include arguments for the test apk, same as above. Note: It is recommended to at least provide the argument for the test filter to save time.

With mb.py, you use -s to specify swarming, and -d to specify dimensions, which narrow down the choice in try bot. The dimensions are added in the form: -d dimension_name dimension_value. You should specify the pool as chromium.tests, the device_os_type as userdebug, and the device_os for whichever Android version you're interested in (e.g. M, N, O, etc). After specifying all your arguments to mb.py, include a --, and after this -- all further arguments are what is passed to the build target (e.g. content_shell_test_apk).

Putting this all together, to run the same tests as above, in the same way, but on the Android M try bots, you would use the command:

tools/mb/mb.py run -s --no-default-dimensions \
    -d pool chromium.tests \
    -d device_os_type userdebug \
    -d device_os M \
    out/Debug \
    content_shell_test_apk \
    -- \
    --num_retries=0 \
    --repeat=-1 \
    --break-on-failure \
    -f "*WebContentsAccessibilityTreeTest*testExample"

Common Android accessibility “gotchas”

  • “name” vs. “text”

    On other platforms, there is a concept of “name”, and throughout the accessibility code there are references to name. In the Android API, this is referred to as “text” and is an attribute in the AccessibilityNodeInfo object. In another platform you may setName() for a node, the equivalent on Android is info.setText(text). In the BrowserAccessibilityAndroid class, the relevant method that provides this information is GetTextContentUTF16.

  • ShouldExposeValueAsName

    The AccessibilityNodeInfo objects of Android do not have a concept of “value”. This makes some strange cases for nodes that have both a value and text or label, and a challenge for how exactly to expose this data through the API. The ShouldExposeValueAsName method of BrowserAccessibilityAndroid returns a boolean of whether or not the node‘s value should be returned for its name (i.e. text, see above). If this value is false, then we concatenate the value with the node’s text and return this from GetTextContentUTF16. In the cases where ShouldExposeValueAsName is true, we expose only the value in the text attribute, and use the “hint” attribute of AccessibilityNodeInfo to expose the rest of the information (text, label, description, placeholder).

  • stateDescription

    The stateDescription attribute of the AccessibilityNodeInfo objects is a recent addition to the API which allows custom text to be added for any node. This text is usually read at the end of a node's announcement, and does not replace any other content and is purely additional information. We make heavy use of the state description to try and capture the richness of the web. For example, the Android API has a concept of checkboxes being checked or unchecked, but it does not have the concept of ‘partially checked’ as we have on the web (kMixed). When a checkbox is partially checked, we include that information in the stateDescription attribute. For some nodes like lists we include a stateDescription of the form “in list, item x of y”. The full list of stateDescriptions can be found in the BrowserAccessibilityAndroid method GetStateDescription.

  • CollectionInfo and CollectionItemInfo

    The AccessibilityNodeInfo object has some child objects that do not always need to be populated, for example CollectionInfo and CollectionItemInfo. Collections work differently on Android than other platforms, namely that a given node does not carry all necessary information to make a determination about the overall collection. As the names might suggest, an item in a collection will have the CollectionItemInfo populated, but not CollectionInfo, whereas the container that holds all the items of the collection will have a CollectionInfo object but not a CollectionItemInfo. When a CollectionItemInfo object is present, it is up to the downstream AT to walk up the tree and gather information about the full Collection. This information is not included on every node.

    These collections are used for any table-like node on Android, such as lists, tables, grids, trees, etc. If a node is not table like or an item of a table, then these child objects would be null. For this example tree, the objects present for each node would be:

    kGenericContainer - CollectionInfo=null; CollectionItemInfo=null
    kList - CollectionInfo=populated; CollectionItemInfo=null
    ++kListItem - CollectionInfo=null; CollectionItemInfo=populated
    ++kListItem - CollectionInfo=null; CollectionItemInfo=populated
    ++kListItem - CollectionInfo=null; CollectionItemInfo=populated
    
  • contentInvalid

    The Android accessibility API has a boolean field isContentInvalid, however this does not play well with downstream AT, so the Chrome code has some special implementation details. The accessibility code reports a page exactly as it is, so if a text field is labeled as contentInvalid, we report the same to all platforms. There are use-cases where a field may be contentInvalid for each character typed until a certain regex is met, e.g. when typing an email, empty or a few characters would be reported as contentInvalid. When isContentInvalid is true on a node's AccessibilityNodeInfo object, then AT (e.g. TalkBack) will proactively announce “Error” or “Content Invalid”, which can be jarring and unexpected for the user. This announcement happens on any change, so every character typed would make this announcement and give a bad user experience. It is the opinion of the Chrome accessibility team that this ought to be fixed by TalkBack, and reporting an invalid node as invalid is the pedantically correct approach. However, in the spirit of giving the best user experience possible, we added two workarounds:

    • The contentInvalid boolean is always false if the number of characters in the field is less than kMinimumCharacterCountForInvalid, currently set to 7. This is done at the BrowserAccessibilityAndroid level.
    • The WebContentsAccessibilityImpl includes a further workaround. contentInvalid will only be reported for a currently focused node, and it will be reported at most once every CONTENT_INVALID_THROTTLE_DELAY seconds, currently set to 4.5s. See the setAccessibilityNodeInfoBooleanAttributes method for the full implementation.
  • isVisibleToUser vs. “offscreen”

    The Android accessibility API includes only one boolean for setting whether or not a node is “visible”, the isVisibleToUser attribute. In general this conflicts with the way the accessibility code treats offscreen. The name isVisibleToUser may suggest that it reflects whether or not the node is currently visible to the user, but a more apt name would be: isPotentiallyVisibleToUser, or isNotProgrammaticallyHidden. Nodes that are scrolled off the screen, and thus not visible to the user, must still report true for isVisibleToUser. The main use-case for this is for AT to allow navigation by element type. For example, if a user wants to navigate by Headings, then an app like TalkBack will only navigate through nodes with a true value for isVisibleToUser. If any node offscreen has isVisibleToUser as false, then it would effectively remove this navigation option. So, the Chrome Android accessibility code reports most nodes as isVisibleToUser, and if the node is actually offscreen (not programmatically hidden but scrolled offscreen), then we include a Bundle extra boolean, “offscreen” so that downstream AT can differentiate between the nodes truly on/off screen.

  • RangeInfo, aria-valuetext, and caching

    The RangeInfo object is another child object of AccessibilityNodeInfo. Unfortunately this object is rather limited in its options, and can only provide float values for a min, max, and current value. There is no concept of a text description, or steps or step size. This clashes with nodes such as sliders with an aria-valuetext, or an indeterminate progress bar, for which we have to add special treatment. As a further complication, AccessibilityEvents also require information on range values when there is a change in value, however the event only allows integer values between 0 and 100 (an integer percentage of the sliders position). BrowserAccessibilityAndroid has a method IsRangeControlWithoutAriaValueText which we use to separate these cases when populating AccessibilityNodeInfo objects and AccessibilityEvents (see web_contents_accessibility_android.cc). Similar to the Collection related objects above, RangeInfo is null for any non-range related nodes.

    This RangeInfo object plays a small role in updating the cached AccessibilityNodeInfo objects above. There is a small bug in the Android framework (which has been fixed on newer versions) which breaks our caching mechanism for range objects. So the UpdateCachedAccessibilityNodeInfo method also updates the RangeInfo object of a node if it has one.

  • Leaf nodes and links

    Android has slightly different IsLeaf logic than other platforms, and this can cause confusion, especially around links. On Android, links are never leafs. See IsLeaf. This is for similar reasons to the isVisibleToUser section above. If a link were a leaf node, and it were to contain something like a Heading, then AT would not be able to traverse that link when navigating by headings because it would only see it is a link. For this reason we always expose the entire child structure of links.

  • Refocusing a node Java-side

    There is a strange bug in Android where objects that are accessibility focused sometimes do not visually update their outline. This does not really block any user flows per se, but we would ideally have the outlines drawn by AT like TalkBack to reflect the correct bounds of the node. There is a simple way to get around this bug, which is to remove focus from the node and refocus it again, which triggers the underlying Android code necessary to update the bounds. In WebContentsAccessibilityImpl we have a method moveAccessibilityFocusToIdAndRefocusIfNeeded which handles this.

  • liveRegions and forced announcements

    There is a boolean for liveRegion in the AccessibilityNodeInfo object that the Chrome accessibility code will never set. This is because TalkBack will read the entirety of the liveRegion node when there is a change. Instead we force a custom announcement with an AccessibilityEvent in the WebContentsAccessibilityImpl's announceLiveRegionText method.

  • sendAccessibilityEvent vs. requestSendAccessibilityEvent

    In the Android framework, there are two methods for sending AccessibiltiyEvents, sendAccessibilityEvent and requestSendAccessibilityEvent. Technically speaking, requestSendAccessibilityEvent will ask the system to send an event, but it doesn't have to send it. For all intents and purposes, we assume that it is always sent, but as a small detail to keep in mind, this is not a guarantee.

  • TYPE_WINDOW_CONTENT_CHANGED events

    The TYPE_WINDOW_CONTENT_CHANGED type AccessibilityEvent is used to tell the framework that something has changed for the given node. We send these events for any change in the web, including scrolling. As a result, this is generally the most frequently sent event, and we can often send too many and put strain on downstream AT. We have included this event as part of our event throttling in the AccessibilityEventDispatcher. We also include a small optimization for a given atomic update. If an atomic update sends more than kMaxContentChangedEventsToFire events (currently set to 5), then any further events are dropped and a single event on the root node is sent instead. This has proven useful for situations such as many nodes being toggled visible at once. See HandleContentChanged in web_contents_accessibility_android.cc.

  • Text selection

    Static text selection exists in Android, but not when a service like TalkBack is enabled. TalkBack allows for text selection inside an editable text field, but not static text inside the web contents. Text selection is also tied to a specific node with a start and end index. This means that we cannot select text across multiple nodes. Ideally the implementation would allow a start and end index on separate nodes, but this work is still in development.

  • Touch exploration

    The way touch exploration works on Android is complicated. The process that happens during any hover event is:

    1. User begins dragging their finger
    2. Java-side View receives a hover event and passes this through to C++
    3. Accessibility code sends a hit test action to the renderer process
    4. The renderer process fires a HOVER accessibility event on the accessibility node at that coordinate
    5. WebContentsAccessibilityImpl#handleHover is called for that node.
    6. We fire a TYPE_VIEW_HOVER_ENTER AccessibilityEvent on that node.
    7. We fire a TYPE_VIEW_HOVER_EXIT AccessibilityEvent on the previous node.
    8. TalkBack sets accessibility focus to the targeted node.
  • WebView and Custom Tabs

    As mentioned at the start of this document, Chrome Android also plays an important role by providing the WebView, which is used by any third party apps that want to include web content in their app. Android also has a unique feature of Custom Tabs, which is a lightweight implementation of Chrome that is somewhere between a WebView and a full browser. The WebView and custom tabs must also be accessible, and so most of this document applies to them as well. Occasionally there will be an edge case or small bug that only happens in WebView, or a feature that needs to be turned off only for WebView/Custom tab (e.g. image descriptions). There is a WebView and Chrome Custom Tabs test app on the chrome-accessibility appspot page, and there are methods on the Java-side that can give signals of whether the current instance is a WebView or Chrome custom tab. Example: isCustomTab.

Recent Progress and Features

The Chrome Android accessibility code continues to evolve each quarter. We have strengthened our testing and the stability of the code, but we also continue to add new features and improvements. Beyond the usual bug fixes, below is a quick summary of some features in the pipeline.

OnDemand AT

We have recently implemented a feature we refer to as “OnDemand AT” for short. This feature is still rolling out and we intend to eventually have it enabled on 100% stable by default. The feature modifies the AccessibilityEventDispatcher that is explained above. If the feature is enabled, then WebContentsAccessibilityImpl will query the Android system to determine the currently enabled accessibility services, as well as the types of data they are interested in, namely the types of AccessibilityEvents they want to know about. When the AccessibilityEventDispatcher is sent an event to add to its queue or dispatch, if that event type is not in the list of AccessibilityEvents relevant to currently enabled accessibility services, the Dispatcher simply drops/ignores the request. Preliminary data shows that this has created a noticeable improvement for accessibility services that do not require the entire suite to function.

AccessibilityPerformanceFiltering

Loosely related to the OnDemand feature above, the “AccessibilityPerformanceFiltering” feature is also a recent addition to improve overall performance. This feature uses the same mechanism as OnDemand to query the currently enabled services and the information they are interested in. AccessibilityPerformanceFiltering then takes this information and uses a different AXMode based on the situation. Each AXMode filters out a different level of information to improve performance while preserving necessary functionality. This effectively does the same thing as OnDemand, but further left/up-the-chain, giving a more significant performance improvement. This feature is still rolling out, and it currently only has three AXModes (basic, form controls only, and complete). As it rolls out and we gather more data we will potentially add more AXModes in the future.

AutoDisableAccessibility

The “AutoDisable” accessibility feature has also been ported to Android. This feature tracks timing between user inputs and accessibility actions to make a determination of whether or not accessibility services are still required. If they are no longer needed by the user, then the accessibility code is disabled. Before this feature, once the code was enabled it would continue to run for the life of the current browser session. This feature is still being rolled out to stable.

Accessibility code in the ClankUI

Most of this document is focused on the accessibility code and work as it relates to the web contents, which is where the Chrome & Chrome OS Accessibility team focuses most of its work. However, some features require a native UI in the browser app, outside the web contents. When these features are added, the line between the accessibility team and the Clank UI team becomes blurred. We traditionally are the owners of this code, but seek regular guidance and approvals from the Clank team as the front-end code must conform to the Clank standards.

AccessibilitySettings

The AccessibilitySettings page is found under the overflow/3-dot menu (Menu>Settings>Accessibility). The page currently contains a slider to change the font scaling of the web contents, options to force enable zoom, show simplified pages, enable image descriptions (see below), and live captions.

The main entry point for this code is here. The code leverages the PreferenceFragment of Android, and so much of the UI and navigation is available out of the box, and the code is relatively simple in that it only needs to respond to user actions/changes and pass this information to the native C++ code.

The settings code is heavily unit tested and stable, so it is rare to have to work in this area.

Image Descriptions

The Clank-side code for the image descriptions feature is a bit more involved. The image descriptions has to track state, determine whether or not to display the option in the overflow menu, show dialogs, and provide toasts to the user. This code is mostly controlled by the ImageDescriptionsController. The image descriptions feature is written using Clank's ‘component’ model, and so almost all the code exists in the directory: chrome/browser/image_descriptions

(The exception being the few hooks that connect this code to the other parts of the Clank UI).

The image descriptions code is heavily unit tested and stable, so it is rare to have to work in this area.

(Upcoming) Page Zoom

An upcoming feature is the page zoom feature, which will allow a more robust way to zoom web contents than the currently existing text scaling of AccessibilitySettings (which will be replaced).

The Clank UI code for this feature has not been developed. More to come.