VVGP features for speaker verification using i-vector framework

Author

de Souza, Cristian Jesus S. ; Gonzalez, Diana Cristina G. ; Lee Luan Ling

Author_Institution

UNICAMP, Campinas, Brazil

fYear

2015

fDate

14-17 June 2015

Firstpage

1

Lastpage

4

Abstract

Spectral parameters alone, especially the Mel-Frequency Cepstral Coefficients (MFCCs) and perceptual linear prediction (PLP) coefficients, have shown good performance in speaker recognition. However, the cepstrum carries only linear information. In this paper, we study the performance of the Variable Variance Gaussian Parameter (VVGP) in a state-of-the-art i-vector speaker verification system. Experimental results on the Ynoguti 2 database indicate that VVGP features is complementary to MFCCs and can improve recognition accuracy.

Keywords

Gaussian processes; cepstral analysis; prediction theory; speaker recognition; vectors; MFCC; Mel-frequency cepstral coefficients; PLP coefficients; VVGP features; Ynoguti 2 database; i-vector framework; i-vector speaker verification system; linear information; perceptual linear prediction coefficients; speaker recognition; spectral parameters; variable variance Gaussian parameter; Feature extraction; Fractals; Geophysical measurements; Mathematical model; Mel frequency cepstral coefficient; Speaker recognition; Speech; Feature extraction; Probabilistic Linear Discriminant Analysis; Speaker Verification; i-vectors;

fLanguage

English

Publisher

ieee

Conference_Titel

Telecommunications (IWT), 2015 International Workshop on

Conference_Location

Santa Rita do Sapucai

Type

conf

DOI

10.1109/IWT.2015.7224579

Filename

7224579