A Modified Speaker Clustering Method for Efficient Speaker Identification

In speaker identification system, along with the growth of the population size, scoring process can be extremely time consuming. In such a case, speaker clustering is generally used to alleviate the situation. K-means is the widely used clustering algorithms, however, its performance suffers from the so-called local optimum problem. To deal with the problem, a novel initialization approach was introduced in this paper, which performs the initialization to the intrinsic spreading patterns of speaker models. In essence, the proposal is of the same spirit to the well-known Canopy mechanism. However, it differs from the Canopy in the aspects of candidate selection and thresholds setting. It is showed, to the application purpose, the proposed approach could work effectively and generates more rational and stable clustering outcome.