Reinhold Weber wrote an interesting article that shows some of the accessibility features of CSS, including aural CSS (the speech media type). However, the sad truth about aural CSS is that it's not supported by the overwhelming majority of screen readers, with few exceptions (as far as I know, Orca for Linux supports it). Why not?
To understand this problem, we have to view the web with the eyes of a screen reader. Screen readers are assistive technologies written for one purpose: helping blind people to use effectively their computer, not only for surfing the web. For that reason, screen readers feature a lot of keyboard shortcuts for helping users to perform tasks that are trivial for other users, such as opening a dialog box, selecting a menu item and so on. When it comes to the web, screen readers implementors soon noticed that a huge amount of web pages were not written with accessibility in mind, because they were full of tables, frames, scripts and other bad practices.
So they decided to protect their users from those accessibility problems by enabling not only keyboard shortcuts, but also other features such as form mode, virtual cursor and virtual memory. In other words, they decided to protect their users from authors because authors seemed not to care about accessibility. Now imagine what would happen if such authors were able to change the aural presentation of a document through CSS. For example, they might specify a background sound played aloud on every single page, thus disturbing the normal reading of the program, or lower the volume of certain elements under a reasonable level and so on. To put it simple: CSS aural is not supported because it poses more accessibility problems than it can actually solve. That's why screen readers don't support it: to prevent authors from overusing such a functionality with predictable side effects.