A speaker model clustering method based on space position

In the speaker recognition system with large numbers of models, the traditional computations of matching one by one can be very time-consuming. In order to solve the problem of fast recognition, this paper proposes a speaker model clustering method based on space position. The idea is: firstly, to divide the training models into multiple layers, then searches class representatives for every layer, then to cluster the training model by the gotten class representatives. This method can greatly reduce the number of models matching, and achieve the goal of fast recognition. The paper takes the Gaussian mixture model (GMM) as the speech model, the experiment shows the recognition average speed of 0.5s per person, and the correct recognition rate less than 1% loss is ensured.