Abstract | ||
---|---|---|
Many methods evaluate the similarity between two vectors x and y by norm or metric learning. They need to get a subtraction vector x - y and then evaluate its length. However, only considering the length of subtraction vector and ignoring its position may lost a lot of information. In this paper, we propose to utilize the position information of subtraction vector to evaluate the similarity. As the subtraction vector between x and y can be expressed either by x - y or by y - x, its distribution is centrosymmetric and redundancy. Thus, only half of the subtraction vectors are chosen and named as subtraction positive vectors. The subtraction positive vectors from different classes or from the same class are then modeled by Gaussian mixture models or deep neural network. Experiments were carried out on speaker verification databases including NIST SRE08, SRE10 and NIST i-vector challenge 2014. Results demonstrate the effectiveness of the proposed method. |
Year | DOI | Venue |
---|---|---|
2019 | 10.1109/APSIPAASC47483.2019.9023082 | Asia-Pacific Signal and Information Processing Association Annual Summit and Conference |
DocType | ISSN | Citations |
Conference | 2309-9402 | 0 |
PageRank | References | Authors |
0.34 | 0 | 4 |
Name | Order | Citations | PageRank |
---|---|---|---|
Liang He | 1 | 67 | 17.35 |
Xianhong Chen | 2 | 1 | 2.04 |
Can Xu | 3 | 4 | 2.14 |
Jia Liu | 4 | 0 | 0.34 |