基于自适应结构图的半监督语音情感特征选择
2021-08-09罗辉韩纪庆
罗辉 韩纪庆



摘 要: 本文研究了语音情感识别中的半监督特征选择问题,即如何利用未标记语音情感数据来帮助选择具有情感判别性的特征。为了解决这个问题,提出了一种新的基于图的半监督特征选择方法。其可以根据标签适应度和流形平滑度,在图上估计一个预测标签矩阵,从而有效地利用标记数据中的标签信息,以及标记数据和未标记数据中的流形结构信息。与现有的基于图的方法相比,该方法能同时进行特征选择和局部结构学习,从而自适应地确定图相似度矩阵。同时,还对图相似度矩阵进行了约束,使其包含更准确的数据结构信息,从而可以选择更有判别性的特征。此外,提出了一种有效的迭代算法来优化该问题。在典型语音情感数据集上的实验结果表明,本文提出的方法是有效的。
关键词: 语音情感识别; 半监督特征选择; 自适应结构图
文章编号: 2095-2163(2021)03-0001-08 中图分类号:TP391.41 文献标志码:A
【Abstract】This paper considers the problem of semi-supervised feature selection in speech emotion recognition, that is, how to use unlabeled speech emotion data to help select the features with emotion discriminability. To address this problem, the paper proposes a novel graph-based semi-supervised feature selection method. The proposed method can estimate a prediction label matrix on the graph with respect to the label fitness and the manifold smoothness, thus it can effectively utilize label information from labeled data as well as a manifold structure information from both labeled and unlabeled data. In comparison with the existing graph-based algorithms, the proposed approach can perform feature selection and local structure learning simultaneously, so the graph similarity matrix can be determined adaptively. At the same time, the paper constrains the similarity matrix to make it contain more accurate data structure information, therefore the proposed approach can select features that are more discriminative. Moreover, an efficient iterative algorithm is proposed to optimize the problem. Experimental results on typical speech emotion datasets show that the proposed method is effective.
【Key words】 speech emotion recognition; semi-supervised feature selection; adaptive structured graph
0 引 言
隨着电子技术和计算机技术的发展,人们需要具有情感识别能力的新型语音对话系统。然而,要实现这一目标,还需要克服许多困难。首先,在特征提取方面,尚不清楚哪些语音特征能有效区分语音情感[1]。其次,不同的句子、说话者、说话风格和语速等因素都会引起不同的声学变化,给语音情感识别增加了新的挑战[1-2]。
特征选择不仅可以突出情感所带来的可变性,还能减少情感之外其它因素的干扰,并能保留原始特征的可解释性[1]。……
