A global covariance matrix based principal component analysis for speaker identification

Authors:
Chagnwoo Seo;Jun Youn;Youngeun Kim;Keunho Sim;Jaekwan Ko;Meihua Zhao;Jongkeun Kim;Heeae Ko;Eunyoung Kim;Younghwan Lim
Affiliations:
The Global School of Media, Soongsil University, Seoul, Korea;The Global School of Media, Soongsil University, Seoul, Korea;The Global School of Media, Soongsil University, Seoul, Korea;The Global School of Media, Soongsil University, Seoul, Korea;The Global School of Media, Soongsil University, Seoul, Korea;The Global School of Media, Soongsil University, Seoul, Korea;The Global School of Media, Soongsil University, Seoul, Korea;The Global School of Media, Soongsil University, Seoul, Korea;The Global School of Media, Soongsil University, Seoul, Korea;The Global School of Media, Soongsil University, Seoul, Korea
Venue:
APCC'09 Proceedings of the 15th Asia-Pacific conference on Communications
Year:
2009

Citing 1
Cited 0

Speaker recognition and speaker normalization by projection to speaker subspace

ICASSP '96 Proceedings of the Acoustics, Speech, and Signal Processing, 1996. on Conference Proceedings., 1996 IEEE International Conference - Volume 01

Quantified Score

Hi-index	0.00

Visualization

Abstract

Principal component analysis (PCA) is an efficient feature extraction method which reduces the dimensions of the feature vectors and removes the correlation among them, with little loss of information, by projecting the original feature space into a small subspace through a transformation. However, it requires a larger amount of training data when calculate the full covariance matrix of each speaker. This paper proposes an efficient global covariance matrix-based PCA for speaker identification. The proposed method uses training data from all speakers to calculate the global covariance matrix and then uses this matrix to find the eigenvalue matrix and the eigenvector matrix to perform PCA. During training and testing, our method uses PCA coefficients that are based on global covariance instead of the PCA coefficients of each speaker. Compared to the conventional PCA and the Gaussian mixture model (GMM) methods, the proposed method shows better performance while requiring less storage space and complexity.