基于深度自编码器子域自适应的跨库语音情感识别

针对不同语料库之间数据分布差异问题,提出一种基于深度自编码器子域自适应的跨库语音情感识别算法.首先,该算法采用两个深度自编码器分别获取源域和目标域表征性强的低维情感特征;然后,利用基于LMMD(local maximum mean discrepancy)的子域自适应模块,实现源域和目标域在不同低维情感类别空间中的特征分布对齐;最后,使用带标签的源域数据进行有监督地训练该模型.在eNTERFACE库为源域、Berlin库为目标域的跨库识别方案中,所提算法的跨库识别准确率相比于其他算法提升了5.26％～19.73％;在Berlin库为源域、eNTERFACE库为目标域的跨库识别方案中,所提算法的跨库识别准确率相比于其他算法提升了7.34％～8.18％.因此,所提方法可以有效地提取不同语料库的共有情感特征并提升了跨库语音情感识别的性能.

Cross-corpus speech emotion recognition based on deep autoencoder subdomain adaptation

To solve the problem of data distribution difference among different corpora, this paper proposed a cross-corpus speech emotion recognition algorithm based on subdomain adaptive deep autoencoder. Firstly, it used two depth autoencoders to obtain representative low-dimensional emotional features of source domain and target domain, respectively. Then, it used an adaptive sub-domain module based on LMMMD to achieve the alignment of feature distribution between source domain and target domain in different low-dimensional emotional category spaces. Finally, it used the tagged source domain data to supervise the training of the model. In the cross-corpus recognition scheme with eNTERFACE library as the source domain and Berlin library as the target domain, the accuracy of the proposed method is 5.26%~19.73% higher than that of the benchmark method. In the cross-corpus recognition scheme with Berlin library as the source domain and eNTERFACE library as the target domain, the accuracy of the proposed method is 7.34%~8.18% higher than that of the benchmark method. Therefore, the proposed method can effectively extract the common sentiment features of different corpora and improve the performance of cross-corpus speech sentiment recognition.