基于短时谱最小均方误差估计的语音增强和剩余噪声衰减

Speech Enhancement Based on Minimum Mean-Square Error Short-Time Spectral Estimation and Residual Noise Reduction

本文研究了三种基于语音短时谱最小均方误差估计的语音增强方法:短时谱幅度最小均方误差估计,短时对数谱最小均方误差估计和短时相对谱幅度最小均方误差估计,在理论分析基础上对它们进行了实验研究.计算机仿真结果表明,在加性白色高斯噪声污染下,当带噪语音信噪比为+5～-10dB时,处理后的语音信噪比提高了3.4～12dB。短时对致谱最小均方误差估计的效果最好,试听实验也证实了这一点。 文中还对增强后剩余噪声的衰减问题进行了研究。利用中心削波并对估值器中的增益函数进行修正,可明显地减弱剩余噪声。

In this paper, we study three types of noisy speech enhancement algorithms based on Minimum Mean-Square Error Short-Time Spectral Estimation: Spectral Amplitude Estimation (MMSE-STSA), LOG-Spectral Amplitude Estimation (LOG-MMSE-STSA). and Relative Spectral Amplitude Estimation (MMSE-REL-STSA). On the basis of theoretical analysis, experimental studies are conducted. Computer simulations show that SNR improvement of 3.4-12dB can be obtained after enhancement processing when the original SNR of noisy speech degraded by Additional White Gaussian Noise ranges from 5dB to-10dB. Among the three algorithms, MMSE-LOG-STSA provides the best SNR improvement, and the SNR results coincide with informal listening.We also study the reduction of residual noise accompanied by the enhanced speech. By using the center clipping and gain function modification of estimator, the residual noise can be reduced significantly.