I can't speak for the speech translator (I think voxforge and CMU sphinx need to get a lot further before we even think about doing any of that kind of stuff).

The picture translation is a different matter. See earlier discussion (yes, it starts out talking AR but a few posts down starts becoming relevant to image translation) and Joaquim's fantastic work on OCRFeeder.

Google's Translate API doesn't look ideal in that it seems geared to embedding in in-page javascript but it might work in a pinch (render the OCR'ed text as HTML and embed the GTrans in that, I suppose).

Of course, you're at the mercy of the OCR engine (I think OCRFeeder uses tesseract? Happy to be corrected) and that is really far from perfect. And if you get unacceptable OCR you'll get garbage from Google translate. So yes, would be nice but I suspect it's early days for this kind of capability. I suppose if OCRFeeder could pipe the output from a particularly well photographed piece of text into an application calling Google's Translate API you might end up with something usable.

That might be an idea for OCRFeeder, make it available as a service for other applications.