Underdetermined blind speech source separation based on deep nearest neighbor clustering algorithm

In this paper, we propose a new autoencoder network architecture with clustering mechanism for underdetermined blind speech source separation, i.e., the number of mixtures is less than that of sources. The autoencoder network is employed to project the mixtures to embedding space and obtain their embedding vectors. The network model additionally incorporates the clustering mechanism and nearest neighbor clustering algorithm to estimate the clustering centers of the embedding vectors. Then, according to the embedding vectors, the hard and the probability assignment method are proposed to assign the mixtures to their corresponding clusters to recover the sources. The experimental results demonstrate that the proposed method yields better performance compared to the baseline algorithms.

Automated summary

In this paper, a novel autoencoder network architecture with clustering mechanism is proposed for underdetermined blind speech source separation. Experimental results demonstrate that the proposed method outperforms baseline algorithms.