Elastic loss for robust speaker verification: A deep learning approach

Speaker verification is a biometric technology that identifies and authenticates individuals based on speech data. It compresses speech features into vector representations and computes similarity scores to determine identity matches. However, intra-speaker variability caused by factors such as emotions, environmental conditions, and context can degrade verification performance. To address this challenge, we integrate elastic loss into ECAPA-TDNN, a deep learning model for speaker feature extraction. Elastic loss utilizes randomly drawn margins from a Gaussian distribution to enhance training robustness and generalization, improving cosine similarity scoring between speech samples. Elastic loss aims to enhance robustness in speaker verification b explicitly handling intra-speaker variability arising from variations in emotional state, environmental noise, and recording conditions. Experimental results on the VoxCeleb dataset demonstrate that our approach outperforms AAM-Softmax in terms of equal error rate(EER), minimum detection cost function (minDCF), and area under the curve (AUC). Additionally, our method shows superior performance across varying speech durations, including segments longer than two seconds.