Speaker segmentation and clustering based on the improved spectral clustering

Efficient speaker segmentation and clustering method based on the improved spectral clustering is proposed in this paper. Traditional speaker segmentation and clustering is performed by the hierarchical clustering algorithms with Bayesian information criterion (BIC) metric and cross likelihood ratio (CLR) metric after the speakers are segmented. Since this method has high computational complexity and may result in a suboptimal solution, we use spectral clustering to overcome this problem and improve the performance of clustering algorithm. First the affinity matrix is constructed with the mean supervector feature transformed by KL kernel mapping. And then the scaling parameter is selected adaptively. The experiments performed on the NIST 1998 multi-speaker corpus show that the proposed method outperforms the baseline system.