基于自编码器的语音情感识别方法研究

Speech emotion recognition based on adversarially trained stacked auto encoder

在语音情感识别的研究中存在特征集维度过高的问题。高维度的特征向量易造成参数过拟合。因此需要一种合适的特征提取与筛选的方法降低特征维度。自编码器是一种应用广泛的特征降维方法,由此本文提出一种基于栈式自编码器,结合对抗训练的方法并在对抗训练中引入Wasserstein距离构造对抗损失函数进行特征降维。实验结果表明,与原始的对抗自编码器相比,经过改进的自编码器在对特征进行同等程度的降维后,准确率平均提高了3.31%。

The feature set dimension is too high in the research of speech emotion recognition.High dimensional eigenvectors tend to cause parameter overfitting.Therefore,an appropriate feature extraction and screening method is needed to reduce the feature dimension.Self-encoder is a kind of characteristic dimension reduction method which is widely used,so this paper proposes a method based on stack self-encoder and combining with confrontation training,and introduces Wasserstein distance to construct confrontation loss function for characteristic dimension reduction.The experimental results show that compared with the original adcs,the accuracy of the improved adcs increases by 3.31%on average after the same degree of dimensionality reduction.