Vocal source features for bilingual speaker identification

Author

Jianglin Wang ; Johnson, Matthew Thomas

Author_Institution

Dept. of Electr. & Comput. Eng., Marquette Univ., Milwaukee, WI, USA

fYear

2013

fDate

6-10 July 2013

Firstpage

170

Lastpage

173

Abstract

This paper introduces the use of two new features for speaker identification, Residual Phase Cepstrum Coefficients (RPCC) and Glottal Flow Cepstrum Coefficients (GLFCC), to capture speaker-specific characteristics from their vocal excitation patterns. Results on a cross-lingual speaker identification task taken from the NIST 2004 SRE demonstrate that these RPCC and GLFCC features are significantly more accurate than traditional mel-frequency cepstral coefficients (MFCC). In particular, these two new features give better results with smaller amounts of training data, due to lower model complexity.

Keywords

Gaussian processes; maximum likelihood estimation; natural language processing; speaker recognition; GLFCC; GMM-UBM; Gaussian mixture model- universal background model; NIST 2004 SRE; RPCC; bilingual speaker identification; cross-lingual speaker identification task; glottal flow cepstrum coefficients; maximum a posteriori adaptation; model complexity; residual phase cepstrum coefficients; speaker-specific characteristics; vocal excitation patterns; vocal source features; Accuracy; Adaptation models; Feature extraction; Filtering; Mel frequency cepstral coefficient; Speaker recognition; Speech; Glottal source excitation; IAIF and GMM; Speaker identification;

fLanguage

English

Publisher

ieee

Conference_Titel

Signal and Information Processing (ChinaSIP), 2013 IEEE China Summit & International Conference on

Conference_Location

Beijing

Type

conf

DOI

10.1109/ChinaSIP.2013.6625321

Filename

6625321