GENERALIZED END-TO-END LOSS FOR SPEAKER VERIFICATION

GENERALIZED END-TO-END LOSS FOR SPEAKER VERIFICATION

9 Nov 2020 | Li Wan, Quan Wang, Alan Papir, Ignacio Lopez Moreno
This paper introduces a new loss function called Generalized End-to-End (GE2E) loss, which enhances the efficiency of training speaker verification models compared to the previous Tuple-Based End-to-End (TE2E) loss function. The GE2E loss function updates the network to emphasize difficult examples at each training step, eliminating the need for an initial example selection stage. This results in a more than 10% reduction in speaker verification Error Rate (EER) and a 60% decrease in training time. The paper also introduces the MultiReader technique, which enables domain adaptation by training a single model to support multiple keywords and dialects. The experimental results demonstrate that the GE2E loss function outperforms the TE2E loss function in both text-dependent and text-independent speaker verification tasks, with significant improvements in accuracy and training efficiency.This paper introduces a new loss function called Generalized End-to-End (GE2E) loss, which enhances the efficiency of training speaker verification models compared to the previous Tuple-Based End-to-End (TE2E) loss function. The GE2E loss function updates the network to emphasize difficult examples at each training step, eliminating the need for an initial example selection stage. This results in a more than 10% reduction in speaker verification Error Rate (EER) and a 60% decrease in training time. The paper also introduces the MultiReader technique, which enables domain adaptation by training a single model to support multiple keywords and dialects. The experimental results demonstrate that the GE2E loss function outperforms the TE2E loss function in both text-dependent and text-independent speaker verification tasks, with significant improvements in accuracy and training efficiency.
Reach us at info@study.space
Understanding Generalized End-to-End Loss for Speaker Verification