Research On The Speaker Recognition Based On Speaker Model Clustering

Speaker recognition is an identity authentication technology, which can automatically realized by using a computer. With its unique advantages, such as convenience, economical efficiency and accuracy etc, it is increasingly becoming a popular mode to identify authentication in people’s daily life, and it has broader application prospect for the future market.The recognition accuracy and robustness were always the research keys of the speaker recognition system, Gaussian Mixture Model (GMM) with its good description to the distribution characteristics, widely used in text-independent speaker recognition. But as the increasing number of the speaker recognition system registration, GMM model recognition needs to take more and more time, the recognition efficiency also gradually become one of the key points to influence the practicability of the system. Based on the Speaker Model Clustering (SMC), this paper proposed a speaker recognition method. This method can ensure the accuracy of the recognition and greatly improve the recognition speed.During the recognition, we could only find a few high-score speakers every time, so what we just need to do are to match the test vectors with these few speaker models and recognize the speaker identity. Based on the similar characteristics of speaker models, SMC clustered the similar speaker models and selected cluster centroids and cluster representatives. In the procedure of testing, we firstly selected the cluster by computing the Euclidean distance between the test vectors and cluster centroids or calculating the logarithmic likelihood between the test vectors and cluster representatives, and then recognized the speaker identity through calculating the logarithmic likelihood between the test vectors and the speaker models which contained in the selected cluster. In case of the possibility of not clustering completely, we chose several higher score clusters to constitute a category to ensure the accuracy of identification. The experimental results show that compared with traditional GMM model, the proposed the method can speed up the recognition speed about four times whit only0.95%loss in recognition accuracy when the cluster number is100and the test class search range is20%. The speaker recognition system’s recognition speed was greatly improved.In order to further improve the recognition speed of the speaker recognition system, this paper also proposed a fusion algorithm between SMC and Pre-Quantization or Pruning. Pre-Quantization or Pruning is to improve recognition speed by dealing with test characteristic vectors, which belongs to the testing stage speed-up techniques; but SMC refers to the stage of gathering trained speaker models into a class after the ending of training process, which belongs to the training stage speed-up techniques. Different stages of the accelerated methods can undertake integration, and further improve the recognition speed of the system.