Sadly matlab still is much easier to use and understand for engineers, which is one of its, if not the, primary market(s).
I prefer using python for machine learning tho. I dislike the libraries written in matlab.
Is numpy+python faster than matlab?
The main reason I say matlab is better for engineers is that its documentation is beyond amazing, if you think of something you need you search in the help browser and there is a good chance it is there with substantial information (something that is much harder with the math/science python libraries) also if you know a functions name you just type: help function and get plenty of information which helps when you forget how to use something and need a quick lookup.
I have done some things with scipy but nothing more than solving some odes and I must say even with all the pains of matlab with writing separate files to solve odes and other such things plus some syntax grievances it is still the easiest solution for me.
I do know that I tried rewriting one of my C++ codes using scipy/numpy and it was about a hundred times slower for that particular task (which did happen to involve very large matrices, approaching the limits of memory).
The earlier solvers took strings with the name of a function whenever they needed a functional argument. Nowadays Matlab provides `function handles' to refer to a function without calling it. The syntax is @function_name, they need this special syntax, because just mentioning function_name itself in your code, calls your function without arguments.
Matlab even support proper closures with lexical scoping, now.
I do agree with all of these advantages of using Matlab in an environment where my data is already cleaned and normalized; I just rarely find myself in that position.
I do not know about I/O libraries. All my Matlab knowledge basically comes from helping out friends with Matlab problems, and reading the manual together with them.
The easiest way to install the suite of scientific computing Python packages on Windows may be PythonXY or Enthought EPD. Under Linux and OSX, users seem to have few problems.