Abstract: Aiming at speech enhancement under non-stationary noise environment and low SNR, an improved spectral subtraction based on real-time noise estimation was proposed. First, voice activity detection was carried out based on selected sub-bands vector distance, and then a real-time adjustment coefficient was defined through noisy speech properties of low-frequency regions and high-frequency regions. The combination of coefficient and voice activity detection could realize the updating of noise estimate.The method allowed tracking the time variation of noise environment.The experiment showed that the method was more effective in reducing background noise, improving SNR and decreasing speech distortion than traditional speech enhancement methods.

Key words: Noise estimation, Voice activity detection, Signal to noise ratio (SNR),Spectral subtraction, Speech enhancement