SUB-BAND PARTITIONING FOR FULL COVARIANCE BASED MISSING DATA SPEAKER RECOGNITION
Missing data processing has been successfully applied to speaker recognition tasks to increase robustness in noise. Previous work has demonstrated the benefit of using full covariance Gaussian mixtures for spectral based marginalization recognition. The drawback of using full covariance modeling with marginalization is the high computational cost. This paper examines subband covariance partitioning as a solution to reduce this computation. The reduced dimensionality of the individual sub-band covariances provides a computational speed-up for the necessary inversion and reliability partitioning operations. Experimental results show that the use of sub-band covariances can produce a significant reduction in the total computation time required to perform bounded marginalization recognition, while still approximating the high performance of full-band modeling.
Speaker recognition Missing data Sub-band partitioning Full covariances
DANIEL PULLELLA MARCO K(U)HNE ROBERTO TOGNERI
School of Electrical, Electronicand Computer Engineering, The University of Western Australia, Crawley, Western Australia 6009
国际会议
The Second International Conference on Information & Systems Sciences(ICISS2008)(第二届信息与系统科学国际会议)
大连
英文
641-648
2008-12-18（万方平台首次上网日期，不代表论文的发表时间）