A Deep Time-delay Embedded Algorithm for Unsupervised Stress Speech Clustering

In this paper, we introduce a new unsupervised clustering algorithm to categorize the stress speech data, called deep time-delay embedded clustering (DTEC). DTEC consisted of a time-delay neural network (TDNN) based autoencoder and the joint supervision of discriminative loss, reconstruction loss, and clustering loss. The TDNN-based autoencoder was designed to transform a high-dimensional input into its low-dimensional output by supervision of reconstruction loss. The inter-cluster push-force and the intracluster pull-force were optimized by supervision of discriminative loss. The distance between the softmax prediction probability and its auxiliary target distribution was minimized by supervision of clustering loss. Based on our experiments, DTEC outperforms the popular deep clustering algorithm and able to increase the clustering performance in terms of accuracy(ACC) and normalized mutual information (NMI).