【摘要】 为解决语音声信号稀疏化表示中稀疏度确定难和稀疏化程度低的问题，提出一种互相关块稀疏贝叶斯学习方法。该方法基于稀疏贝叶斯学习理论，利用语音声信号的块稀疏性和时间相关性特征，构建了语音声信号稀疏解的互相关块稀疏空间结构，使其在特征空间内展现更充分的稀疏性。该方法可有效提高语音声信号的稀疏表示。经数值仿真验证构建的声特征惩罚矩阵D对稀疏解x的约束作用显著提高，x中相邻的数据块密切相关，单一的噪声峰值被抑制，说明能得到最优稀疏解，并且提高了稀疏化程度。更多还原

【Abstract】 To solve the problems of difficult sparsity determination and low sparsity in the sparse representation of speech sound signals,a mutual correlation block sparse learning method is proposed. The method is based on sparse Bayesian learning theory and uses the block sparsity and time-dependent characteristics of speech sound signals to construct a mutual correlation block sparse space structure for the sparse solution of speech sound signals,so that they exhibit more adequate sparsity in the feature space. The method can effectively improve the sparse representation of speech acoustic signals. The constructed acoustic feature penalty matrix D is verified by numerical simulation to significantly improve the constraint effect of the sparse solution x. The numerical simulation verifies that the constructed acoustic feature penalty matrix D significantly improves the constraint on the sparse solution x. The adjacent data blocks in x are closely related and a single noise peak is suppressed,indicating that the optimal sparse solution can be obtained and that the degree of sparsification is improved.更多还原