A Novel Hybrid Neural Embedding Extractor for Text Independent Speaker Verification

Speaker embedding extraction is crucial for neural automatic speaker verification systems. In this work, we propose a novel hybrid neural embedding framework that integrates frequency- and channel-wise Selective Kernel Attention (SKA) into a 2D-CNN-based feature extraction module. This integration improves the aggregation of frequency-channel information, enhancing the extraction of discriminative speaker embeddings. The feature extraction module is connected to a frame-level network, combining a Time-Delay Neural Network (TDNN)-Long Short-Term Memory (LSTM) hybrid with a fully TDNN network in a cascaded structure. To capture speaker information at the utterance level, we use Multi-Level Attentive Statistics Pooling (MLASP), which incorporates local statistics and leverages the complementarity of different networks. MLASP also helps integrate previously overlooked features, enhancing the robustness of the learned embeddings. The entire framework is trained with the additive angular margin softmax (AAMSoftmax) objective, creating an embedding space where embeddings of the same speaker are close together, and those of different speakers are well separated. Experimental results on the VoxCeleb and CNCeleb corpora demonstrate that our approach outperforms both baseline and state-of-the-art speaker verification systems trained on the same datasets.