We investigated the effects of gender, nationalityand emotion information on SV by using MTL andDAT. Four effective systems were proposed by combiningspeaker information and different types of domaininformation in NN training stage. More specifically,MTL-based method was used to enhance thelearning of gender and nationality information inMTG, MTN and MTGN systems. The informationlearning of different emotions of a certain speakerwas suppressed using DAT-based method in EDATsystem. Finally, a linear scoring fusion method wasemployed to combine the advantages of differentsystems. The results indicate that enhance genderand nationality information learning by using MTLbasedmethods can significantly improve the performanceof SV. The results also indicate that the effectof emotion information is suppressed by usingDAT-based method also beneficial for SV. Finally,compared with a baseline system, the performanceof our systems achieved 16.4% and 22.9% relativeimprovements in the EER of MTL and DAT-basedsystems, respectively. Moreover, the relationship ofdifferent speech information can also be referencedto improve the recognition performance in other researchfields such as SER and nationality recognition.
展开▼