GMM-UBM for text-dependent speaker recognition
Wanli Chen, Qingyang Hong, Ximin Li · 2012
Traditional Text-Dependent Speaker Recognition (TDSR) systems model the user-specific spoken passwords with frame-based features such as mel frequency cepstral coefficient (MFCC) and use Dynamic Time Warping (DTW) or hidden Markov Model (HMM) classifiers to handle the variable length of the feature vector sequence. However, DTW can't deal with cross-channel issue while HMM needs more computational complexity and storage space. In this paper, we introduce text-independent framework GMM-UBM into text-dependent field. It not only solves intersession problem but also a compromise between model accuracy and computational cost. Moreover, a more accurate UBM will get lower EER. A new UBM initialization method, LBG-VQ-EM, will be proposed. Experiments shows that it is better than conventional initialization way like K-means. And we also compare the performance of GMM-UBM and DTW, and two stacked methods of training utterances: frame-based and wave-based. The experimental results showed the performance of GMM-UBM exceeded that of DTW, and that of frame-based outperformed that of wave-based.