The Auditory Characteristics And Noise Estimation In Speech Enhancement Algorithm Research

As the inevitability of the presence of noise caused a sharp deterioration phenomenon in voice system performance, and thus directly affect the quality of voice reconstruction, leading to a rapid decline of voice quality. However, how as much as possible remove noise and extracting the pure voice signal in the noisy speech signal is currently a hot topic in the field of speech signal processing.This paper is a research of the speech enhancement, based on spectral subtraction, this research proposed one method based on the auditory characteristics and noise estimation that apply to noise reduction. This method used a more excellent noise estimation algorithm in low signal-to-noise ratio (SNR) non-stationary environment input combined the auditory masking effect of people to realize speech enhancement. So this method greatly reduced the impact of residual’musical noise’, and also reduced the enhanced signal distortion degree, so it improved the quality of enhanced voice.The minimum control recursive average (MCRA) method is a kind of new and effective noise estimation method. It does not need the voice/silent decision of speech, so it more close to the soft decision method than the traditional voice activity detection (VAD) of noise estimation method. And this method recalculate the noise power spectrum for each frame, then it makes noise spectrum can be constantly updated in all of the frame, realizing real-time tracking noise changing needs. Therefore, MCRA method is a very good noise estimation method in non-stationary. The noise estimated is close to the original noise; then after spectral subtraction, the enhanced signal is more’pure’, so is suitable for real time realization.The experimental results show that in the high input SNR stationary noise environment, the method of this paper is relative with based on auditory masking effect algorithm, but in low SNR non-stationary noise input, the enhanced signal of this method is obviously well than the enhanced signal of traditional method. This method improves the output SNR and reduces the time domain distortion degree of enhanced signal, so improves the quality of the output voice. Therefore, the proposed speech enhancement algorithm inhibited’musical noise’ at the same time, also enhanced the hearing effect.