The subject of extracting multiple speech signals from a single mixed recording, which is referred to single channel speech separation, has received considerable attention in recent years and many model-based techniques have been proposed. A major problem of most of these systems is their inability to deal with the situation in which the signals are combined at different levels of energies because they assume that the data used in the test and training phase have equal levels of energies, where, this assumption hardly occurs in reality. Our proposed method based on MIXMAX approximation and sub-section vector quantization (VQ) is an attempt to overcome this limitation. The proposed technique is compared with a technique in which a gain adapted minimum mean square error estimator is derived to estimate the separated signals. Through experiments we show that our proposed method outperforms this method in terms of SNR results and also reduces computational complexity.