Automata learning techniques are getting significant importance for their applications in a wide variety of software engineering problems, especially in the analysis and testing of complex systems. In recent studies, a previous learning approach [1] has been extended to synthesize Mealy machine models which are specifically tailored for I/O based systems. In this paper, we discuss the inference of Mealy machines and propose improvements that reduces the worst-time learning complexity of the existing algorithm. The gain over the complexity of the proposed algorithm has also been confirmed by experimentation on a large set of finite state machines.