An improved GMM-based clustering algorithm for efficient speaker identification

In large population speaker identification (SI) system, likelihood computations during testing stage can be time-consuming. In such a case, clustering method is applied to this situation. But the traditional clustering algorithm based on K-means is sensitive to the randomly chosen initial cluster centers. To address this issue, the paper proposes an improved clustering algorithm which uses an initial clustering method to choose the cluster centroids and utilizes T-test metrics as the distance measure. In fact, the proposed initial clustering method is of the same essence to the subtractive clustering algorithm. However, it differs from the subtractive clustering in two aspects of candidate selection and cluster radius. According to the experimental result, the improved GMM-based clustering algorithm shows the better performance both on recognition rate and accuracy of clustering, compared to the conventional clustering algorithm.