Rapid Speaker Recognition Based On GMM-UBM

Due to its flexibility and facilitation in application, text-independent speaker recognition system has become a hot topic in the field of speech recognition. Since NIST (National Institute of Standards and Technology) 1999 sets the Gaussian Mixture Model - Universal Background Model (GMM-UBM) as a reference system to obtain excellent recognition rate, it has functioned as a baseline system for this research area, and got better and better through improvement. Although the speaker recognition system has achieved a relatively satisfactory result, it requires much time for the calculation of the likelihood before matches, which makes the system recognition speed decline significantly. As a result, its practicality in application is not that promising. The main goal of this paper is to achieve better and quicker speaker recognition without affecting the accuracy of recognition rate.As to the large calculation, slow speed in speaker recognition process, this paper makes improvement based on tree structure of the selection algorithm, top-down searching UBM in the output test speech feature vector core of the likelihood distribution of the highest points . It saves effort when matching with the target speaker model since it is only necessary to calculate the core of the likelihood distribution. Reference system with the improved algorithm the core selected speed has increased to 14.7 times. Because the order of feature vector having no effect to the results, so after combined with the vector sequence rearrangements pruning algorithm, the system speed increased by 21.7 times, with recognition rate slightly lower. To improve the recognition rate, the paper introduces the kernel function in SVM (Support Vector Machine) to the speaker recognition, which basically shares the same recognition rate with the reference system.For the open set recognition problems, this paper proposes the concept of probability threshold to resolve in the open set female voice recognition rate significantly decreased. The recognition speed would decline after using probability threshold and the recognition rate increased by 0.7%. This paper improves to compute the confidence algorithm on the segment to solve some of the match problems on the final recognition results, and different parameters would be used to carry out experiments, the final choice the average of each short segment as its confidence, and then a three layer feed forward network use to make up confidence. Compared to the reference system using the likelihood ratio, the correct rate of inside increased by 2.6%, and the error rate of outside fell by 2%.