Thank you for your suggestions and the link to that great article. I understand it a little bit better now, but I'm still somewhat confused. If screenreader users also get the keyword and comment information, wouldn't anything repeated in the description be redundant?

F.e., I prefer the description "Sam from Life on Mars looking confused", but that can also be deduced from the keyword and comment. Wouldn't it then be better to literally describe what is in the icon and say "Man's face reflected in a windshield"?

If I do this at all I want to do it right :)