Preliminary Study on Self-contained UBM Construction for Speaker Recognition

摘要

Although speaker recognition technology has evolved into some new stages recently, GMM-UBM (Gaussian Mixture Model-Universal Background Model) has always been the base module for the newly developed methods such as SVM, JFA and i-vector. Because of its simplicity, flexibility and robustness, GMM-UBM has been used as a benchmark system for research reference. For traditional UBM construction, speech data from a lot of speakers other than the target speakers should be obtained, which means much cost of data collection. In this paper, we make preliminary exploration on a new approach to train the UBM, named as self-contained UBM, in which only the target speakers' training data were used. We study several strategies of speaker selection for the self-contained UBM construction, gradually reduced from 50 to 3 speakers. Experiments on MASC@CCNT show that our self-contained UBM obtain considerable recognition rate compared with traditional UBM, while needing far less training data thus less training time. Furthermore, we find out that the obtained good ternary UBM speakers have an interesting characteristic of spanning a triangle (UBM speaker triangle) after dimension reduction of MFCC features with PCA.