An important problem in signal processing is the "cocktail party problem", where several people are speaking at the same time and the objective is to separate the different speakers, typically using several microphones placed in different localities. Numerous techniques had been proposed to solve the cocktail party problem, with various degrees of success. Many of these techniques work very well for artificially mixed speech signals, but when it comes to real recordings, even with two speakers, the success is often mixed. In this talk, we present a very robust method for solving the cocktail party problem in real recording with two speakers based on time-frequency separation.