基于高斯-拉普拉斯-伽玛模型和人耳听觉掩蔽效应的信号子空间语音增强算法

针对信号子空间语音增强算法中的子空间选择和线性滤波器中噪声功率谱和拉格朗日乘子的估计问题,用高斯、拉普拉斯和伽玛模型描述了语音的分布,提出了利用目标语音概率最大化来确定信号子空间维度的方法。在噪声子空间上,利用条件概率估计出噪声功率谱。接着,为了合理地折中增强语音中的残余噪声和语音畸变,提出了一种基于人耳听觉掩蔽效应的拉格朗日乘子估计方法。实验证明,在多项语音质量评价指标上,所提算法都取得了更好的结果。所提的信号子空间算法比传统的信号子空间算法更有效地消除了噪声,使得恢复的语音具有更好的质量。 

A subspace speech enhancement algorithm based on Gaussian-Laplacian-Gamma statistical models and masking properties of human ears

Subject to the subspace selection problem and the noise and Lagrange multipliers estimation problems,the Gaussian,Laplacian and Gamma models are used to describe the distribution of the speech.A subspace selection method is proposed by maximizing the speech presence probability.In the noise subspace,the conditional probability is used to estimate the noise power spectrum.To balance the noise reduction and speech distortion,Lagrange multipliers are appropriately estimated based on the masking properties of human ears.Experiments show that,the proposed subspace method produces impressive results in terms of quality measures of the enhanced speech.The proposed algorithm more effectively reduces the noise and makes the enhanced speech have better qualities.