Despite the growing popularity of metric learning approaches, very little\nwork has attempted to perform a fair comparison of these techniques for speaker\nverification. We try to fill this gap and compare several metric learning loss\nfunctions in a systematic manner on the VoxCeleb dataset. The first family of\nloss functions is derived from the cross entropy loss (usually used for\nsupervised classification) and includes the congenerous cosine loss, the\nadditive angular margin loss, and the center loss. The second family of loss\nfunctions focuses on the similarity between training samples and includes the\ncontrastive loss and the triplet loss. We show that the additive angular margin\nloss function outperforms all other loss functions in the study, while learning\nmore robust representations. Based on a combination of SincNet trainable\nfeatures and the x-vector architecture, the network used in this paper brings\nus a step closer to a really-end-to-end speaker verification system, when\ncombined with the additive angular margin loss, while still being competitive\nwith the x-vector baseline. In the spirit of reproducible research, we also\nrelease open source Python code for reproducing our results, and share\npretrained PyTorch models on torch.hub that can be used either directly or\nafter fine-tuning.\n