Svm-based Speaker The Rejection Policy Research

In the speaker recognition system,speaker detection and rejection is very important. The speaker's voices are fuzzy or the speakers would probably not be in the training set, this could result in incorrect recognition,that could affect the recognition rate of system.In this paper, we do a brief introduction to speaker recognition system, including the background of speaker recognition, a framework of speaker recognition, the verification contents of the speaker and its development. We established and implemented an online speaker verification system, and completed speaker voice pre-emphasis, feature extraction, Gaussian mixture model training and recognition, rejection strategy of speaker recognition.We used adaptive GMM as speaker recognition method, namely we used the same one's voice to amend the original template parameters. We proposed a new feature for the speaker rejection. That means we used the recognition results of the candidates, the normalized score and the normalized first-order difference as the feature inputting support vector machine classification. According to classification results, we can take the rapid to identification. So this method which combine Gaussian and support vector machine achieved the rapid rejection for impostor. This method is a direct result of the use of identifying information extraction features of the GMM. It needs a smaller amount of storage required and calculates simply and effectively. So it is suitable for speaker recognition system rejection.