Enough screen reader to catch yourself

You are not going to become a screen reader user. Somebody who uses NVDA every day operates it at a speed and with a mental model you will not acquire on a Tuesday evening, and any tutorial implying otherwise is selling something. That gap is the actual reason developers skip this step: the honest version of the task looks enormous, so it never gets started, and the component ships having been checked by nobody.

The task is smaller than that. You are not evaluating whether the interface is good. You are checking whether it says what you think it says. Does the approve control announce as a button. Does it have a name. Does the confidence chip get announced at all. Every one of those is a yes-or-no answer available in a handful of keystrokes, and every one of them is a mistake you made and can fix.

The size of the ask, stated by the people teaching it

WebAIM’s own framing of its NVDA article is that it is “designed to help new or novice NVDA users learn how to test web content”, and it is candid about the ramp: “Plan on spending some time (perhaps a few hours) becoming comfortable using NVDA.” A few hours, once, for a capability you keep. WebAIM is a nonprofit that also sells accessibility training, which is worth naming whenever you quote them, and it cuts against their commercial interest to tell you the ramp is short.

The two to learn are NVDA on Windows and VoiceOver on macOS, for a plain reason: both are free, and between them they cover the two machines you already have. That reason is deliberately not a market-share argument. WebAIM’s Screen Reader User Survey #10 collected 1,539 valid responses in December 2023 and January 2024 and reports JAWS at 40.5%, NVDA at 37.7% and VoiceOver at 9.7% as primary screen readers. Those numbers get quoted as market share constantly, and WebAIM says plainly that they are not: “The sample was not controlled and may not represent all screen reader users.” A self-selected survey is useful context and cannot settle which screen readers your product must support. Cite it for what it is or do not cite it.

NVDA: the commands that do the work

Install it, and during setup choose CapsLock as the NVDA modifier key, which is what WebAIM recommends and what makes every command below reachable one-handed. The commands, from the article:

Learn that last one first. In browse mode, single-letter keys are navigation commands, which is how H jumps between headings. In focus mode, keystrokes pass through to the browser so you can type into a field. WebAIM notes that NVDA switches automatically in most cases and that NVDA + Space toggles the modes manually. Almost every confusing first session with a screen reader is the mode being the opposite of what the tester assumed. If your custom combobox swallows a letter key and nothing happens, check the mode before you file a bug.

VoiceOver: the same job, a different modifier

VoiceOver is already on the Mac. Start it with Command + F5, adding fn on laptops where the function row needs it. The modifier is Control + Option, written VO everywhere in the documentation.

• VO + A starts reading; Control alone stops it.
• VO + → and VO + ← step through items one at a time. This is the one you will live in.
• VO + Command + H jumps to the next heading; swap H for L for links, J for form controls, T for tables.
• VO + Space activates the link or control you are on.
• VO + U opens the rotor, which is VoiceOver’s equivalent of the Elements List.
• Shift + VO + I reads webpage statistics, which is a fast sanity check on whether the page has any headings at all.

One caveat matters more than any command, because it changes what your test result means. WebAIM: “VoiceOver currently functions best with the Safari web browser.” A screen reader is not a property of a component. It is half of a pair with a browser, and the pair is what behaves. “It works with VoiceOver” is not a claim. “VoiceOver with Safari on macOS, tested on this date” is.

The four checks worth doing every time

This is the loop. It takes about five minutes per component once the keystrokes are in your fingers, which is the whole argument for spending the few hours.

1. Tab through and listen to the names. Every control should announce a name, a role, and its state where it has one. A button that announces as “button” with no name is 4.1.2 Name, Role, Value, Level A, heard rather than inferred. An icon button whose name is the icon’s file name is the same failure wearing a disguise.
2. Press H repeatedly. You are checking that the heading structure exists and matches the visual hierarchy. A panel that looks like a section and is built from a bold div is silent here, which is 1.3.1 Info and Relationships, Level A.
3. Open the Elements List or the rotor. This shows you the page the way a screen reader user actually navigates it, as a list of links, headings and buttons stripped of layout. If your component is incomprehensible in that view, it is incomprehensible in use.
4. Operate the widget without looking. Turn the monitor off or close your eyes. Open the field picker, choose an option, confirm the choice was announced. This is the step that finds the missing announcement, because your eyes stop covering for it.

Check four is also where the APG’s own warning gets satisfied. That guide says testing assistive technology interoperability is essential before shipping code from it, and the four patterns you built in the previous lesson have not been listened to yet.

What this does not cover

This module ends with four patterns built against the guide, a focus contract you can test in ninety seconds, and a conformance note whose supported set is finally something you stood behind. All of it applies to interfaces that hold still.

Point any of it at a token stream and it comes apart. Every keystroke here assumes content that is there when you go looking for it, and a streaming response is content that changes while it is being read. The announcing-the-stream lesson opens the next module on exactly that, and it opens on an empty shelf: the W3C’s own editor’s draft on the accessibility of machine learning and generative AI, checked again while this course was written, has nothing to say about announcing incrementally generated text. The pattern has to be built out of 4.1.3 Status Messages, aria-busy, and the Understanding document’s own warning about being too chatty, and the course says where that reasoning came from rather than dressing it up as a citation.