Note: chrome://keyboard works great with select-to-speak, but the embedded keyboard does not!

It does seem like the virtual keyboard is in the a11y tree though:

>> desktop.findAll({attributes: {'name': 'caps lock'}})
This returns a static text and an inlineTextBox in the right location: {left: 204, top: 141, width: 73, height: 20}

However they are labeled as offscreen, and they are in a window marked as 'invisible'.

>> capsLockNodes[0].parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.state
{invisible: true}
>> capsLockNodes[0].parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.role
"window"
>> capsLockNodes[0].parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.parent.location
{left: 0, top: 488, width: 1366, height: 280}

The keyboard is in a position:fixed container, which might be a potential cause of the issue? Although the same is true in chrome://keyboard.

Possibly relevant:

Seems like most Accessibility events are not getting fired when the keyboard pops up, only when it goes back down again.

For example, if you close the main chrome browser window (to avoid extra messages) and pull up the keyboard, there are no accessibility events logged by render_accessibility_impl.cc. But when you close the keyboard, lots of events are logged. 

 Issue 830126  has been merged into this issue.

Select to Speak works on the keyboard now. However, node bounds are incorrect (although they ARE correct in ChromeVox!?) That will be tracked separately.