Multi-Feature Fusion Using Multi-GMM Supervector for SVM Speaker Verification

This paper proposes a novel multi-feature fusion approach using Multi-GMM supervector and Support Vector Machine for text-independent speaker verification. By the UBMMAP framework, the variable number of feature vectors (MFCC, LPCC) can be transformed into a vector (GMM supervector). Concatenating the GMM supervectors from different features, a new Multi-GMM supervector is formed for SVM. Experiments on text-independent speaker verification in NIST´04 10sec-10sec female data showed the successful fusion of MFCC and LPCC in feature level.