Studies On Speaker Recognition Based On SVM And GMM

Speaker recognition is a process of identifying the corresponding speaker, according to the parameters which represent the physiological and behavioral characteristics of the speaker's voice. As a biometric authentication technology, it is an important research direction of the speech signal processing, with a wide range of applications, prompting more people to study it.At present, the speaker recognition based on the closed-set has been made relatively good progress. But the recognition performance of the speaker recognition based on an open-set needs to be improved. The open-set and the closed-set is a partition according to testing sets of speaker. When the testing set of speakers is a subset of the training set, it is called the speaker recognition based on the close-set. When the testing set is not restricted by the training set, no matter if training or not, it is called the speaker recognition based on the open-set. The speaker recognition based on the close-set is different from the speaker recognition based on the open-set. Besides the different testing set, the key origin of the difference is the method of recognition. There is two part of the recognition based on the open-set. One is that judging the testing speaker is whether a member of training set, another is identifying the testing speaker is which speaker in training set.This thesis is committed to the speaker recognition based on the open-set and proposes a new recognition method, which is based on SVM and GMM. The speaker distinguish and recognition contains two parts, one is the speaker distinguish and another is the speaker recognition. The speaker distinguish is to distinguish the testing speaker is whether a member of training set. If it is true, the speaker recognition needs to be carried out. The Implementation is based on SVM-GMM model, which is a mixed model of the support vector machine (SVM) and the Gaussian mixture model (GMM).Support vector machine model is based on statistical learning theory, VC dimension and structural risk minimization theory. On condition of limited sample information, the model aims of finding the best compromise between complexity and learning ability to obtain a good generalization. A large number of experiments have confirmed that the SVM with an excellent classification ability. Because of this, the model has been selected as the first step in its rough classification model. The Ga-ussian mixture model is a linear combination of multiple Gaussian, fitting the characteristic distribution characteristic distribution of the speaker. It is good at expressing the internal similarity of the speaker voice characteristics. The previous studies have shown that the model has good performance in the speaker verification system. Therefore, the Gaussian mixture model is the second sophisticated model.In the recognition phase, firstly, the testing speaker is classified by support vector machine model firstly. Then, the result of classification is verified by Gaussian mixture model, determining whether the testing speaker is its classification in the first step. If so, it indicates that the test speaker is assigned to categories corresponding to their speaker;if not, the testing speaker is a speaker outside of training set. This study was to achieve the distinguished speaker in this way.Experiments show that the proposed method in the thesis is effective and can improve the rate of distinguishing the speaker of outside of training set effectively.