基于深度神经网络和区分性联合字典的无监督单通道盲源分离技术研究

The cross projection engenders when mixed speech signal is represented over joint dictionary because of the bad distinguishing ability of joint dictionary in single-channel blind source separation (SCBSS) using sparse decomposition theory, which leads to bad separation performance. In addition, in actual speech separation tasks, it is often not known which gender combination of mixed speech comes from. Therefore, an unsupervised SCBSS technology based on deep neural network and discriminative joint dictionary is researched. This technology detects the gender combination of mixed speech based on the deep neural network model, and then selects the corresponding discriminatory joint dictionary for separation, which can achieve high-quality unsupervised SCBSS. The main research contents of this project are shown as follows. (1)Based on DNN bottleneck feature fusion and CNN deep and shallow feature fusion, a deep neural network model for mixed speech gender detection is constructed. (2) Based on the autocorrelation of speech signals and OMP template matching criteria, a quasi-LP matrix is constructed. (3) According to the atomic correlations of different identity sub-dictionaries, common sub-dictionary containing the common characteristics of speech signals is constructed. (4) Based on the similarities and differences of source signals, a new optimization function is constructed to learn the discriminative joint dictionary. (5) Based on the characteristics of the speech signal, a reconstruction algorithm is constructed by smoothing function to reconstruct the speech signal with high quality.

基于稀疏分解实现单通道盲源分离时，由于传统方法构建的联合字典区分能力弱，当混合语音信号在联合字典上表示时会产生“交叉投影”，导致分离性能较差。另外，实际语音分离任务中，往往不知道混合语音是来自哪种性别组合的语音。因此，本课题研究基于深度神经网络和区分性联合字典的无监督单通道盲源分离技术。该技术基于深度神经网络模型进行混合语音性别组合检测，然后选取相应的区分性联合字典进行分离，以实现高质量的无监督单通道盲源分离。研究内容如下：(1)基于DNN瓶颈层特征融合、CNN深浅度特征融合等来构造混合语音性别检测模型；(2)基于语音自相关特性和基于OMP模板匹配准则构造准LP矩阵；(3)根据不同身份子字典的原子相关性，构造含有语音信号公共特性的子字典；(4)基于不同源信号的共性和差异性，构造新的优化函数学习得到区分性联合字典；(5)从语音信号本身特性出发，构造平滑函数和使语音信号高质量重构的重构算法。

本课题针对混合语音信号分离时会产生“交叉投影”导致语音分离性能较差，以及实际语音分离任务中往往不知道混合语音是来自全男性、全女性、异性别组合中哪种性别组合的问题，从构造混合语音性别检测模型、特征提取融合方案、深度神经网络结构、稀疏矩阵、含有语音信号公共特性的子字典、区分性联合字典、高质量重构、深度神经网络双输出联合约束、多重联合约束以及时空特征融合技术等方面展开，重点研究了混合语音性别检测的深度神经网络模型，以及基于深度神经网络和区分性联合字典的无监督单通道盲源分离技术。本课题提出一种基于共振峰分布的改进MFCC参数提取算法、一种基于深度瓶颈特征与声学特征寻优融合的分类特征集构造方法、一种基于深浅层高斯超矢量最优融合的特征提取方案、基于DNN的两阶段瓶颈特征提取和多分类器联合判决算法方案、一种基于双支路卷积神经网络-支持向量机的性别组合检测模型，来判断混合语音的两个说话人是男-男、男-女还是女-女组合。提出基于区分性约束和抗混淆约束的优化函数用于学习双层字典，实现了单通道的语音和噪声的分离。提出一种用于解决语音信号在联合字典上的“交叉投影”问题的联合约束字典学习方法，来提升单通道语音与噪声的分离效果。提出一种基于深度神经网络双输出联合约束的损失函数，实现了多种性别组合的单通道语音分离。提出一种基于卷积神经网络掩码与幅度谱联合优化的单通道语音分离方法、一种基于整合优化器的单声道语音分离方法和一种基于双输出神经网络的多重联合约束损失函数的单通道语音分离方法。提出一种强约束字典优化函数, 针对相似信号混合实现了基于强约束字典和深度神经网络的高质量两阶段单通道语音分离。