Angular Softmax Loss for End-to-end Speaker Verification
Угловая функция потерь Softmax для сквозной верификации говорящего
2018-11-01
SCID: 54.1/fzfr2v2b
Discuss with AI
Fisher datasetangular softmax lossend-to-end speaker verificationprobabilistic linear discriminant analysisspeaker embeddings
Figures from the paper
Abstract (AI)
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.
Key Findings
1
A-softmax incorporates inter-class angular margins, producing more discriminative speaker embeddings than softmax and triplet losses while remaining easy and stable to use.
2
Angular softmax (A-softmax) is introduced as a feature-learning loss for end-to-end speaker verification.
3
Combining an A-softmax-trained front end with PLDA back-end scoring further improves performance under short enrollment and test utterance conditions.
4
Using A-softmax in end-to-end training achieves significant equal error rate reductions on part of the Fisher dataset.
Research Object
end-to-end speaker verification systems
Research Subject
discriminative speaker-embedding learning and verification performance, including equal error rate and short-utterance scoring, using angular softmax loss with PLDA back-end scoring
Publication Details
Publication Date
2018-11-01
Journal
Publisher
ISSN
Cited by
48
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest
References available in scid.ai6
FaceNet: A unified embedding for face recognition and clustering2015
Kaldi Speech Recognition Toolkit2024
Front-End Factor Analysis for Speaker Verification2010
Generalized End-to-End Loss for Speaker Verification2018
End-to-end text-dependent speaker verification2016
Deep neural network-based speaker embeddings for end-to-end speaker verification2016