Text-independent speaker identification based on feature transformation to phoneme-independent subspace

In text-independent (TI) speaker identification, the variation of phonetic information strongly affects the performance of speaker identification. If this phonetic information in his/her speech data can be suppressed, a robust TI speaker identification system will be realized by using speech features having less phonetic information. In this paper, we propose a TI speaker identification method that suppresses the phonetic information by a subspace method, under the assumption that a subspace with large variance in the speech feature space is a ldquophoneme-dependent subspacerdquo and a complementary subspace of it is a ldquophoneme-independent subspacerdquo. Principal Component Analysis (PCA) is utilized to construct these subspaces. We carried out GMM-based speaker identification experiments using both a new feature vector of the proposed method and the conventional MFCC. As a result, the proposed method reduced the identification error rate by 21% compared with the conventional MFCC.