Discriminative Vector Learning with Application To Single Channel Speech Separation

In this paper, we introduce a discriminative vector learning method. First, the points are transferred into the discriminative vector using two backbone networks. These vectors are easily separated by a simple clustering algorithm. Among them, the similarity of vectors in different clusters is low, and the similarity of vectors in the same cluster is high. This property is very important in image segmentation, audio separation, and data clustering problems. In our work, we apply our discriminative vector learning method for single-microphone speech separation, casting this task as spectrogram segmentation. Overall, the experiments show that our method greatly improves the performance compared to other deep clustering methods for speech separation.