An Abnormal Speech Detection Algorithm Based on GMM-UBM

To overcome the defects of common used algorithms based on model for abnormal speech recognition, which existed insufficient training data and difficult to fit each type of abnormal characters, an abnormal speech detection method based on GMM-UBM was proposed in this paper. For compensating the defects of methods based on model which difficult to deal with the diversification speech. Firstly, many normal utterances and unknowing type abnormal utterances came from different speaker, were used to train the GMM-UBM for normal speech and abnormal speech, respectively; secondly, the GMM-UBM obtained by training normal speech and abnormal speech were used to s core for these testing utterances. From the results show that compared with GMM and GMM-SVM methods under 24 Gaussians and the ratio of training speech and testing is 6:4, the correct classification ratio of this proposed have 6.1% and 4.4% improvement, respectively.