Efficient speaker identification using speaker model clustering

In large population speaker identification (SI) systems, likelihood computations between an unknown speaker's feature set and the registered speaker models can be very time-consuming and impose a bottleneck. For applications requiring fast SI, this is a problem. In prior work, we proposed the use of clusters of speaker models so that during the test stage, only a small proportion of speaker models in selected clusters are used in the likelihood computations resulting in a speedup of 2× without loss in accuracy. In this paper, we improve the method by incorporating log-likelihoods into the initial clustering as well as cluster selection. The new method allows for fewer clusters to be searched and thus higher speed-up factors while still maintaining acceptable accuracy levels.