Article ID Journal Published Year Pages File Type
4977796 Speech Communication 2017 16 Pages PDF
Abstract
Conventional speaker verification systems utilize information from different feature representations by means of fusion. In this paper, we propose an alternative technique which achieves a similar effect but utilizes a more effective feature selection technique. The underlying assumption of the method is that different speakers may be better represented, and hence better verified, in different feature spaces. This technique, which we term as feature-switching, performs verification using a feature representation most suitable to the speaker under consideration. Out of a possible set of candidate representations, the most optimal representation for a speaker is determined during enrollment. Then verification is performed using the optimal feature of the claimed speaker. Experimental evaluation of feature-switching is performed utilizing the classical GMM-UBM speaker verification system, as well as the i-vector-based verification system. Our results show that feature-switching achieves improved performance compared to conventional as well as fusion-based systems.
Related Topics
Physical Sciences and Engineering Computer Science Signal Processing
Authors
, , ,