Screen readers do not obey language changes in HTML

I have a rather general-purpose question aimed for Screen reader and Accessibility API specialists. The question category is internationalisation.
WCAG 2.0 requires the semantics for the main document human language, as well as for the language of document parts if it differs from the main document language. There is no word in the guidelines that screen readers may ignore the specified language in so-called focus mode.
Currently neither NVDA nor JAWS (appologize for my Windows-related approach) obey language switching if in focus / forms / application mode. We have in many languages worldwide in the last time more and more anglicisms and in Germany / Austria in order to get a very good user experience and legally needed Accessibility certification according to BITV or WCAG2.0, your website and especially web app should work with Screen readers optimally. The current Screen readers switch synth languages only in browse / virtual cursor mode - so to say only in the document text. If you implement a single-page application putting role=application to the body element because your web app has only GUI widgets, and you declare the complete document as held in  English language by putting lang=en to the html element, JAWS and NVDA will speak the contents of your widgets further on in the first user language - in the language of the Windows system - let's say in German. It does not help to remove the role=application: by tabbing into your widgets (tablist, tree, form inputs, listbox items etc.) the screen readers automatically switch to focus mode (thats OK so far), but then they stop recognizing the document language. The same with partial language changes put with span elements directly to a word or a phrase within the contents.
Now my question: Is this problem due to MSAA / UIAutomation / iAccessible2 missing the possibility for the Browser to set proper language for AT? Has the iAccessible2 maybe also this problem or the Screen readers do not use it? Have Screen readers technically a possibility to combine their focus mode with their awareness of the defined human language for the whole document and for its parts that they show during function in the browse / virtual cursor mode?
Thanks in advance.
Mario