In this paper, we propose a new shift-invariant feature extraction method based on tensor analysis for robust text-independent speaker Identification. Multiple factors including time, frequency, scale and phase are investigated to extract the essential features of speech signal. In order to explore the shift-invariant characteristic of spectral features, we propose a convolutive model to learn the spectral basis function in tensor structure. Nonnegative assumption is also imposed on the CANDECOMP model to ensure the sparsity and preserve robust features. Experimental results demonstrate that our proposed method can improve the recognition accuracy specifically in noise conditions.
|Number of pages||8|
|Journal||Journal of Computational Information Systems|
|Publication status||Published - 1 Nov 2012|
- Convolutive tensor basis function
- Robust feature extraction
- Sparse representation
- Speaker identification