Speaker identification in a multimodal interface

The article presents the development of a speaker identification system as one part of the multimodal interface for the HBB-NEXT project. A short introduction to a speaker identification problem in the context of HBB-NEXT project is given. Then we focus on the design, optimization and method selection process in order to realize a real time, text independent speaker identification application, namely: selection and optimization of acoustical features, and selection and optimization of cooperating classification methods. All adjustments and evaluations were executed on a speaker database containing testing and training sections. The highest accuracies reaching 95% were observed for frequency ranges 300-9000Hz, MFCC and KNN methods.