There are some loss functions we can used in speaker recognition task, they are:
- AAM-Softmax
- DAM-Softmax
- AM-Softmax
- Triplet Loss
- GE2E Loss
- Angular Prototypical
et al.
However, which loss function can get the best performance in speaker recognition?
In paper: In defence of metric learning for speaker recognition, we can find this answer.
From this paper, we can find Angular Prototypical Loss will get the best performance.
Here is the comparative results.