Powerful Speaker Embedding Training Framework by Adversarially Disentangled Identity Representation
The main challenge of speaker verification in the wild is the interference caused by irrelevant information in speech and the lack of speaker labels in speech datasets. In order to solve the above problems, we propose a novel speaker embedding training framework based on adversarially disentangled identity representation. Our key insight is to adversarially learn the identity-purified features for speaker verification, and learn an identity-irrelated feature whose speaker information cannot be distinguished. Based on the existing state-of-the-art speaker verification models, we improve them without adjusting the structure and hyper-parameters of any model. Experiments prove that the framework we propose can significantly improve the performance of speaker verification from the original model without any empirical adjustments. Proving that it is particularly useful for alleviating the lack of speaker labels.
PDF Abstract
VoxCeleb1
