Effective Modeling of Critical Contextual Information for TDNN-based Speaker Verification

Abstract

Today, Time Delay Neural Network (TDNN) has become the mainstream architecture for speaker verification task, in which the ECAPA-TDNN is one of the state-of-the-art models. The current works that focus on improving TDNN primarily address the limitations of TDNN in modeling global information and bridge the gap between TDNN and 2-Dimensional convolutions. However, the hierarchical convolutional structure in the SE-Res2Block proposed by ECAPA-TDNN cannot make full use of the contextual information, resulting in the weak ability of ECAPA-TDNN to model effective context dependencies. To this end, three improved architectures based on ECAPA-TDNN are proposed to fully and effectively extract multi-scale features with context dependence and then aggregate these features. The experimental results on VoxCeleb and CN-Celeb verify the effectiveness of the three proposed architectures. One of these architectures achieves nearly a 23% lower Equal Error Rate compared to that of ECAPA-TDNN on VoxCeleb1-O dataset, demonstrating the competitive performance achievable among the current TDNN architectures under the comparable parameter count.