自适应高斯混合模型语音增强方法

语音增强是解决噪声污染的有效方法，它的首要目标是在接收端尽可能从带噪语音中恢复纯净的语音信号．针对噪声环境下的语音增强问题，提出了一种语音增强新方法．该方法利用小波子带的方向性特点以及小波系数尺度内的相关性，将小波系数的概率分布建模为一种自适应高斯混合模型，在贝叶斯框架中采用这种概率模型可以得到一种具有空间自适应性的贝叶斯萎缩函数．利用这种萎缩函数可以实现对小波系数的修正．仿真实验表明，该算法对于噪声有较好的抑制作用，该算法在主观和客观测试中都具有良好的语音增强效果，可以在语音识别、语音编码中获得应用．

A speech enhancement method based on adaptive Gaussian mixture model

Speech enhancement is an efficient method in removing noises. The aim of speech enhancement is resuming the pure speech signal from the speech signals containing noise to a great extent. In order to solve the problem of speech enhancement in noise environment, a new speech enhancement method was proposed. This method models the distribution of wavelet coefficients as an adaptive Gaussian mixture model. This model takes into account intra scale dependencies between wavelet coefficients. Based on this model in a Bayesian framework, a spatially adaptive Bayesian shrinkage function was obtained and each modified coefficient was decided separately. Simulation results show this algorithm is effective in reducing the noise; this algorithm possesses good performance both in objective and subjective tests, so it can be used in speech recognition and speech coding.