Obviously, this depends on individual learners used. But since there are no iterative loops in the architecture, our time will only be the sum of the time required to complete each component of the architecture.
Testing time is significantly less than training time. Firstly, usually only a small subset of features needs to be extracted; rather than all possible features. Secondly, no clustering is done as the data arrives. Thus we can see that this architecture is biased towards speed of recognition, rather than training.