I definitely think image recognition has improved a lot, both in speed
and accuracy. However, even a difference like 50 milliseconds may be
noticeable by an experienced screen reader user, especially if one
uses speech at 400 words per minute or more. This is one of the
reasons why many blind users (including myself) still prefer the
text-mode console over a graphical terminal for command line work. The
graphical terminal is certainly very usable and works well for some
scenarios,, but there is quite a noticeable difference in performance
when using the text-mode console.
Regards,
Rynhardt
On Sat, May 29, 2021 at 9:48 AM Matan Safriel wrote:
>
> Of course, the approach whereby the application developers interleave
> semantics about the content beyond the particular way it is laid out in 2D,
> as you mention, is very sensible and robust as long as GUI toolkits and
> development processes enable it as a default piece of development, and
> application developers and designers put that extra semantic information in,
> such as pointing out that a grid actually bears plain list semantics and no
> special significance to it being a grid in a particular case.
>
> Do bear in mind that the potential for robust text reading from images has
> significantly improved since the time that the current accessibility paradigm
> took over, so I am not sure I see a reason it would be "always slower" and
> things like that ― these affirmations are probably not true today ― if one
> embarked on a machine learning project on this.
>
> On Sat, May 29, 2021 at 9:52 AM Rynhardt Kruger via gnome-accessibility-list
> wrote:
>>
>> I agree with you, it would be a useful fallback. It would never be a primary
>> solution though, it is essentially screen-scraping, and would have the same
>> disadvantages as screen-scraping approaches that were used before
>> accessibility API:
>>
>>
>> * Accessibility APIs make it the app developer's responsibility to implement
>> propper accessibility, this is by design. App developers know not just their
>> app, but also the content associated with it, and therefore can implement an
>> accessible experience that may be different from the visual layout and yet
>> more efficient for AT users. An example is the list of recommendations on
>> Youtube. Visually they are in a grid, but the screen reader sees it as a
>> list of headings.
>> * One of the most important requirements for a screen reader is
>> responsiveness. The quicker a blind user knows about an update in the
>> interface, the better. Even a slight delay before an announcement may result
>> in an interface that feels sluggish. This is why blind users often use
>> speech at a very fast speaking rate. A pattern-based approach will always be
>> slower than just reading the state via the accessibility API.
>> * It is dependent on the visual layout, which means accessibility may break
>> just because an app got new icons, or a few components shifted position.
>> These things are independent from the accessibility API.
>>
>>
>> Still, it may be a useful approach to consider for special edge cases, and I
>> am interested to see what happens in this space.
>>
>>
>> Regards,
>>
>>
>> Rynhardt
>>
>>
>> On Fri, 28 May 2021, 20:10 Matan Safriel via gnome-accessibility-list,
>> wrote:
>>>
>>> Hi Shadyar,
>>>
>>> Not an immediate solution at all, but I would say that AI (Machine
>>> Learning) which snapshots the screen or window and is able to extract the
>>> text from the snapshot image to then read it aloud, might be superior to
>>> legacy accessibility API paradigms which rely on the application developers
>>> to interleave "accessibility" (ARIA etc.) information in each and every
>>> field.
>>>
>>> Or at least as an augmentation that should be able to provide a really
>>> great fallback to any ARIA-like paradigm.
>>>
>>> It would be a project, sure, but it is really very accomplishable at this
>>> time and age.
>>>
>>> Hopefully one day our desktops will be more fluid than only providing voice
>>> services on top a graphical interaction interface, but a lot can be done
>>> till then by leveraging Computer Vision AI in this space. Sorry again that
>>> this is not any immediate solution.
>>>
>>> Matan
>>>
>>> On Fri, May 28, 2021 at 4:04 AM Shadyar Khodayari via
>>> gnome-accessibility-list wrote:
>>>>
>>>> Hello
>>>> I'm a blind computer engineer, a developer and f
