All technologies involved here are error-prone. Overcoming their limitation will warrant some kind of reinvention of the wheel that will be impressive to behold if it works.

Simply employing current Speech-to-Text output in conjunction with machine translation alone won’t cut it (even with automatic speech recognition accuracy rates well in the eighties.) Translating clean text into another language is a noisy channel problem and while there are machine translation approaches that work with “broken” input, there is no wholesale solution that systematically models speech recognition errors. The problem then becomes semantic one (not statistics, which Google exceeds at).

In any case, Google’s solution may just blow our minds.

Google buying Nuance I suspect is unlikely, not only because of company value, but also as the latter consists of a whole range of product and professional service groups (beyond the obvious speech technology data/expertise) that Google will not be interested in, but which will be hard to break out of the overall structure.