Speaker verification based on SVDD
Yuhu Zhou, Yong Gong, Jinming Wang, Jian Sun, Xiongwei Zhang, Tao Zhu · 2010 3rd International Congress on Image and Signal Processing · 2010
A new speaker verification method based on SVDD is proposed from another viewpoint to solve the problem such that the recognition rate of the speaker verification system based on GMM is not very high. The traditional hard decision method is also changed, instead a new soft decision means based on the sample acceptance rate is introduced, whose main advantage is normalized the 'probability' or 'confidence' to the value (0,1) and the threshold is comparative steady so as to be easy setting. Experiments are based on TIMIT database, selecting 50 people, with cross validation method, taking EER for the system performance evaluation criteria, comparing SVDD with GMM systems in unified threshold or independent threshold. Experiments result show that this method obviously improves the recognition accuracy and robustness; the running time of the system based on SVDD reduced 26.2% and EER decreased 22.5%; furthermore, when under the conditions of insufficient training samples, SVDD gain more obvious advantage over GMM in teams of training time, running time and recognition accuracy.