Angular Softmax Loss for End-to-end Speaker Verification

TL;DR

This paper introduces the use of angular softmax loss in end-to-end speaker verification, significantly improving discriminative feature learning and system performance, especially under short utterance conditions.

Contribution

The paper pioneers applying A-softmax loss to speaker verification and demonstrates its effectiveness in enhancing discriminative features and overall system accuracy.

Findings

A-softmax loss reduces EER in speaker verification.

Combining A-softmax with PLDA scoring improves performance on short utterances.

Experiments on Fisher dataset show significant accuracy gains.

Abstract

End-to-end speaker verification systems have received increasing interests. The traditional i-vector approach trains a generative model (basically a factor-analysis model) to extract i-vectors as speaker embeddings. In contrast, the end-to-end approach directly trains a discriminative model (often a neural network) to learn discriminative speaker embeddings; a crucial component is the training criterion. In this paper, we use angular softmax (A-softmax), which is originally proposed for face verification, as the loss function for feature learning in end-to-end speaker verification. By introducing margins between classes into softmax loss, A-softmax can learn more discriminative features than softmax loss and triplet loss, and at the same time, is easy and stable for usage. We make two contributions in this work. 1) We introduce A-softmax loss into end-to-end speaker verification and achieve significant EER reductions. 2) We find that the combination of using A-softmax in training the front-end and using PLDA in the back-end scoring further boosts the performance of end-to-end systems under short utterance condition (short in both enrollment and test). Experiments are conducted on part of $Fisher$ dataset and demonstrate the improvements of using A-softmax.