That is what we asked ourselves when we built our first recognizer and spoken term detection system for Tamil, the surprise language of the OpenKWS 2014 evaluation. It seems the recognition should be good - but it is not, we’re confusing lots of words, and keyword spotting is also hard. We’re assuming that we’re no worse than anybody else, and we are not (see the OpenKWS 2014 Workshop) - but why?

http://en.wikipedia.org/wiki/Tamil_language