Abstract: Aiming at non-stationary noise under low signal-to-noise ratio environment situation,the geometric approach to spectral subtraction algorithm produces a larger voice distortion shortcomings,An improved algorithm is proposed based on auditory masking effect,thus to control the residual noise strength under the masking threshold.By calculating the PESQ、oSNR、Spectrogram and time domain simulation waveform diagram analysis and subjective listening test show that improved geometric spectral subtraction algorithm not only improves the signal-to-noise ratio of speech signals,but also reduces the residual noise and speech distortion,to improve the speech intelligibility.