Sort GMM based Speaker Verification Research
|School||University of Science and Technology of China|
|Course||Circuits and Systems|
|Keywords||Text-independent speaker verification Universal Background Model Gaussian mixture model Sort Gaussian mixture model Search width|
With the progress of society and civilization continues to develop, require specific identity of the speaker identification requirements are also increasing. In all authentication technologies, biometric authentication technology because of its own based on human physiology and behavior and other characteristics of the unique advantages of the widely used, showing the biometric authentication technology in the practical application of broad prospects. The existing biometric authentication technology, and text-independent speaker verification is considered to be the most natural biometric authentication technologies, it is through a specific speaker's voice speaker authentication, but also in speech recognition research a very important research direction. Most text-independent speaker verification systems are based on short-term cepstrum parameters and GMM-UBM-MAP model structure, the use of this structure and the text-independent speaker verification systems have reached a very high recognition rate. System performance and computational identification speaker verification system is to select the two most important criteria. UBM in traditional training process, for each input feature vector, to calculate all the UBM Likelihood Gaussian components, the UBM Gaussian mixture model of higher order, and is made by a number of different voice training impersonate obtained So training UBM considerable amount of computation, which to some extent limit the GMM-UBM-MAP-based speaker verification system structure applications. For text-independent speaker verification problem, this paper delves into reducing the amount of computation UBM training, improve training the UBM speeds. The main contents are as follows: 1. Described in detail based GMM-UBM-MAP structure and text-independent speaker verification system, discussed the GMM training algorithm and the MAP algorithm. 2 introduces the distinction of having a good SVM model, in-depth discussion of SVM speaker verification system applied problems, and compares GMM-UBM-MAP structure and GMM-Sup-SVM structure and the text independent speaker Verify that the system's performance. 3 introduces two short-time analysis based channel cepstrum MFCC, LPCC extraction methods, and discuss their use in speaker verification effectiveness and robustness. 4 In-depth analysis of the UBM model training process large amount of computation problems, introduce a Gaussian mixture model based on the sort of training the UBM method, which can reduce the amount of computation UBM training, improve training UBM speed, it will the UBM the degree of mixing of the various criteria according to a predetermined order, the input speech frame only need training to participate in all parts of the degree of mixing in the Gaussian component of training, thereby reducing the amount of computation UBM training. Sort Gaussian mixture model approach using UBM training, not only reduces the amount of computation UBM training, and almost does not affect the recognition performance.