Output details

Learning speaker-specific characteristics with a deep neural architecture

<24> A substantial challenge in speech information processing is the existence of different yet entangled information components that interfere with each other in tasks ranging from speech to speaker recognition. This work is significant, because the previous work can extract speaker-specific features for only known speakers (close-set), whereas this approach learns generic speaker-specific characteristics, disentangling different information components and then extracting speaker-specific information for any speakers (open set). Empirical evaluations reveal significant performance benefits in speaker recognition.