基于层次稀疏DBN的瓶颈特征提取方法∗

摘要

To overcome the drawbacks of original speech features that long temporal speeches and the supervised information can not be effectively utilized and the training time cost is high, a bottleneck feature extraction method based on hierarchical deep sparse belief network is presented. The overlapping group lasso is used as the sparse regularization constraint of the objective function of deep belief network to obtain a deep sparse belief network with a higher speed. To make full use of the hierarchical structure, two sparse deep belief networks are connected in series to enhance the discriminant ability of the bottleneck features. The experimental results on phoneme recognition task show that the proposed feature is effective.%针对现有语音特征无法有效利用长时段语音和监督性类别信息，及现有瓶颈特征提取方法耗时过长等缺陷，提出基于层次结构稀疏深度可信神经网络的瓶颈特征提取方法。该方法将重叠组套索作为深度可信神经网络目标函数的稀疏正则项使用，从而构建训练速度更快的稀疏深度可信神经网络。然后利用层次结构的网络架构方式，将两个稀疏深度可信神经网络串联后使用，进一步增强瓶颈特征的判决能力。文中将此瓶颈特征应用于音素识别中，实验表明该特征的有效性。