Detail publikace

Abstrakt

The task of speaker recognition may be viewed as a validation process, where a decision about the true identity of an unknown speaker represented by her/his speech recording has to be made. The focus will be laid on the Text Independent Speaker Recognition (TISR). Hence, none a-priori assumption is made about the presence of acoustic events (phones, syllables, words, etc.) occurring in the speech recording. At first Cepstral Coefficients (CCs) are extracted from the speech recording (an acoustic space is formed), and subsequently, GMMs are trained to represent the speaker specic regions in the acoustic space. To cope with small amount of training data adaptation of an Universal Background Model (UBM) is utlilized. Since ML estimation relies only on target data, it does not reflect the topology/location/characteristic of impostor data, it is quite handy to involve also discriminative techniques providing such an additional information - in this paper Support Vector Machines (SVMs) are investigated. Main focus is laid on an approach combining GMMs and SVMs with additional improvements.