A Universal Phoneme-Set Based Language Independent Short Utterance Speaker Recognition

In the field of speaker recognition, short utterance speaker recognition (SUSR) has been attracting more and more attention in recent years. Despite the advancement in this technology and use of phonetic cues for speaker recognition, the role of individual phonemes in carrying speaker information is yet quite an open issue. This paper presents a novel idea of using phoneme classes as a basis for SUSR. For the present work, we have restricted ourselves to vowel classes and defined combined vowel classes in two languages, i.e. English and Chinese. These sets are used to develop the universal background phoneme-class model (UBPM) and then for training and testing over conventional GMM-UBM systems. Experimental results have proved that speech segments, as short as phonemes, are surprisingly important areas that carry useful speaker information.