Cosine metric learning based speaker verification

• Published in: Speech communication Vol. 118; pp. 10 - 20
• Main Authors: Bai, Zhongxin, Zhang, Xiao-Lei, Chen, Jingdong
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.04.2020; Elsevier Science Ltd
• Subjects: Algorithms; Compensation; Cosine metric learning; Discriminant analysis; Inter-session variability compensation; Learning; Machine learning; Regularization; Speaker identification; Speaker verification; Speech recognition; Variance; Verification; Speaker verification; Cosine metric learning; Inter-session variability compensation
• ISSN: 0167-6393; 1872-7182
• DOI: 10.1016/j.specom.2020.02.003

•It proposes two cosine metric learning (CML) back-end algorithms. The first one, named m-CML, aims to enlarge the between-class distance with a regularization term to control the within-class variance. The second one, named v-CML, attempts to reduce the within-class variance with a regularization term to prevent the between-class distance from getting smaller.•It combines m-CML with an i-vector front-end since m-CML is good at enlarging the between-class distance of Gaussian score distributions.•It combines v-CML with a d-vector or x-vector front-end as v-CML is able to reduce the within-class variance of heavy-tailed score distributions significantly.•Experimental results on the NIST and SITW speaker recognition evaluation corpora with both i-vector, d-vector and x-vector front-ends demonstrate the effectiveness of the proposed algorithms. The performance of speaker verification depends on the overlap region of the decision scores of true and imposter trials. Motivated by the fact that the overlap region can be reduced by maximizing the between-class distance while minimizing the within-class variance of the trials, we present in this paper two cosine metric learning (CML) back-end algorithms. The first one, named m-CML, aims to enlarge the between-class distance with a regularization term to control the within-class variance. The second one, named v-CML, attempts to reduce the within-class variance with a regularization term to prevent the between-class distance from getting smaller. The regularization terms in the CML methods can be initialized by a traditional channel compensation method, e.g., the linear discriminant analysis. These two algorithms are combined with front-end processing for speaker verification. To validate their effectiveness, m-CML is combined with an i-vector front-end since it is good at enlarging the between-class distance of Gaussian score distributions while v-CML is combined with a d-vector or x-vector front-end as it is able to reduce the within-class variance of heavy-tailed score distributions significantly. Experimental results on the NIST and SITW speaker recognition evaluation corpora show that the proposed algorithms outperform their initialization channel compensation methods, and are competitive to the probabilistic linear discriminant analysis back-end in terms of performance. For comparison, we also applied the m-CML and v-CML methods to the i-vector and x-vector front-ends.