Within- and Between-Class Sample Interpolation Based Supervised Metric Learning for Speaker Verification

Metric learning aims to pull together the samples belonging to the same class and push apart those from different classes in embedding space. Existing methods may suffer from inadequate and low-quality sample pairs, resulting unsatisfactory speaker verification (SV) performance. To address this issue, we propose the data augmentation methods in the embedding space to guarantee sufficient and high-quality negative points for metric learning, termed as within-class and between-class points interpolation generation (WBIG). Furthermore, the strategy of hard negative pair mining (HDPM) is also considered in WBIG. It is shown that WBIG is simple and flexible enough to be incorporated into existing metric learning method, such as supervised contrastive loss (SCL). Experiments on CNCeleb and VoxCeleb demonstrate the superiority of WBIG, and achieve relative performance improvement in terms of EER by 9.74% and 9.95% compared to the baseline system, separately.