Speaker Identification And Verification Based On Cepstral Features And Fuzzy Nonlinear Classif

Abstract

This paper presents an application of fuzzy nonlinear classifier to speaker identification and verification. This classifier is closely related to a modification of a classical Takagi-Sugeno-Kang inference system and is based on a fuzzy moving consequents in If-Then rules. Since fuzzy classifiers based on structural risk minimization have not been applied in speaker recognition area so far, this work presents a novel approach to the problem of speaker modeling. Although voice biometrics is dominated by statistical approach like Gaussian mixture models GMM's and vector quantization VQ, other speaker recognition methods are demanded, especially those which do not require a lot of training utterances and have good generalization properties. All research is based on Polish speech corpus ROBOT designed for testing speech algorithms. Provided results show that fuzzy approach may have similar or even better performance than standard methods