基于非负矩阵分解的语音深层低维特征提取方法

发布时间：2018-06-08 01:52

本文选题：连续语音识别 + 深层神经网络　；参考：《数据采集与处理》2017年05期

【摘要】：作为一种基于深层神经网络提取的低维特征,瓶颈特征在连续语音识别中取得了很大的成功。然而训练瓶颈结构的深层神经网络时,瓶颈层的存在会降低网络输出层的帧准确率,进而反过来影响该特征的性能。针对这一问题,本文基于非负矩阵分解算法,提出一种利用不包含瓶颈层的深层神经网络提取低维特征的方法。该方法利用半非负矩阵分解和凸非负矩阵分解算法对隐含层权值矩阵分解得到基矩阵,将其作为新的特征层权值矩阵,然后在该层不设置偏移向量的情况下,通过数据前向传播提取新型特征。实验表明,该特征具有较为稳定的规律,且适用于不同的识别任务和网络结构。当使用训练数据充足的语料进行实验时,该特征表现出同瓶颈特征几乎相同的识别性能;而在低资源环境下,基于该特征识别系统的识别率明显优于深层神经网络混合识别系统和瓶颈特征识别系统。
[Abstract]:As a low-dimensional feature based on deep neural network, bottleneck feature has been successfully used in continuous speech recognition. However, when training the deep neural network with the bottleneck structure, the existence of the bottleneck layer will reduce the frame accuracy of the network output layer, which in turn will affect the performance of the feature. In order to solve this problem, based on the non-negative matrix decomposition algorithm, this paper proposes a method to extract low-dimensional features by using deep-seated neural networks without bottleneck layer. The method uses semi-nonnegative matrix decomposition and convex non-negative matrix decomposition algorithm to decompose the hidden layer weight matrix to obtain the base matrix, which is regarded as a new characteristic layer weight matrix. New features are extracted by data forward propagation. The experimental results show that this feature has relatively stable rules and is suitable for different recognition tasks and network structures. When experimenting with language data with sufficient training data, the feature exhibits almost the same recognition performance as the bottleneck feature, while in low-resource environments, the performance of the feature is similar to that of the bottleneck feature. The recognition rate of the system based on this feature is obviously better than that of the hybrid recognition system based on deep neural network and the bottleneck feature recognition system.
【作者单位】：解放军信息工程大学信息系统工程学院;
【基金】：国家自然科学基金(61175017,61403415)资助项目
【分类号】：TN912.34;TP183

【相似文献】