In the current system for the reconstruction of speech, it is used iterative reconstruction framework of short time Fourier transformation (STFT). When a lot of noise is added to the input speech, the iterative reconstruction framework of STFT can not properly reconstruct the input speech. In this paper, we offer a method that by using it, we will be able to extract the original signal from the noisy input signal. For achieving this goal, in each step of the iterative reconstruction of input speech, by taking a threshold value For magnitude spectrum, we remove the smaller amounts of threshold and for preserving the original signal features, we consider the phase spectrum then with combination of the modified magnitude spectrum information and the phase spectrum information, we reconstruction part of the signal in each iteration. Two experiments were examined. In the first experiment, we evaluated the reconstruction of input speech with changing the threshold value and in the second experiment, we evaluated the reconstruction of input speech with different numbers of iterations. The results showed that by using our method, when a lot of noise is added to the input speech, we can reconstruct the original signal very well.