Well, depends on what you want to do. Good far field speech recognition requires on the order of 10k hours of training data to build a good acoustic model for

Well, depends on what you want to do. Good far field speech recognition requires on the order of 10k hours of training data to build a good acoustic model for large vocabulary problems. And it needs to be recorded with the same mic set up as you expect at run time. I am continually frustrated by people that think speech recognition is a software problem.