In order to eliminate the influence of power instability on the signal detection in laser speech transmission system, a method combining ensemble empirical mode decomposition (EEMD) and correlation coefficient is proposed to process the signal. The method decomposes the signal into several intrinsic mode functions through EEMD, and calculates the correlation coefficient between each component and the original signal. Then a determination threshold is set to distinguish the real signal and IMF component of trend term. It can overcome such shortcomings existing in traditional methods as the modal aliasing and need for subjective judgment of trend components. Simulation and experimental results show that the noise suppression effect of the improved EEMD is better than wavelet denoising. Furthermore, this method can effectively remove the trend caused by linear change of laser power and sinusoidal change of frequency less than 15Hz.