SE-TDN-LSTM speaker recognition algorithm based on residual connection

To address the issues of low performance and poor robustness of TDNN based on complex conditions in capturing contextual feature information and processing noise, the SE-TDNN-LSTM speaker recognition model based on residual connection is proposed. The model first utilizes the SE-TDNN module to extract local features of speech signals, and enhances their local importance to suppress the influence of redundant information such as noise; Secondly, the residual connection is applied for feature information sharing, and the learned multi-scale features are transmitted to the LSTM-TDNN module to capture long-term dependencies and extract higher-level features, to complete the modeling of contextual global information; Finally, frame-level feature aggregation is performed through a statistical pooling layer with an attention mechanism to obtain speaker features with higher discriminability. The experimental results on the unconstrained Chinese speech dataset of CN-Celeb show that the SE-TDNN-LSTM speaker recognition model based on residual connection reduces the equal error rate (EER) and minimum normalized detection cost function(minDCF) by 2.80% and 0.135 respectively compared to the X-vectors model, achieving an EER of 11.41% and a minDCF of 0.494, and has high accuracy and robustness.