Speaker Recognition with Spectral Dimension Features of Human Voices for Personal Authentication

Biometric recognition is more and more important due to security applications all over the world. Mobile phone becomes popular in recent years. Therefore, voice recognition on mobile devices for recognizing a speaker’s identity plays a potential role. This paper presents a speaker recognition method which combines a non-linear feature, named spectral dimension (SD), with Mel Frequency Cepstral Coefficients (MFCC). In order to improve the performance of the proposed scheme, the Mel-scale method is adopted for allocating sub-bands and the pattern matching is trained by Gaussian mixture model. Some problems related to spectral dimension are discussed and the comparison with other simple spectral features is made. We observe that our proposed methods can improve the performance in different components. For instance, speaker verification combining MFCC with our proposed SD features gives a good performance of EER=2.31% by 32_Multi-GMM. The relative improvement of about 22% may be achieved, which is better than the method that is based only on MFCC with EER=2.96%. Index Terms – Biometric Recognition, Personal Authentication, Speaker Identification, Speaker Verification, Fractal Dimension, Spectral Dimension.