Speaker Verification Based on Tide-Ripple Convolution Neural Network

ObjectiveState-of-the-art speaker verification models typically rely on fixed receptive fields, which limits their ability to represent multi-scale acoustic patterns while increasing parameter counts and computational loads. Speech contains layered temporal–spectral structures, yet the use of dynamic receptive fields to characterize these structures is still not well explored. The design principles for effective dynamic receptive field mechanisms also remain unclear.MethodsInspired by the non-linear coupling behavior of tidal surges, a Tide-Ripple Convolution (TR-Conv) layer is proposed to form a more effective receptive field. TR-Conv constructs primary and auxiliary receptive fields within a window by applying power-of-two interpolation. It then employs a scan-pooling mechanism to capture salient information outside the window and an operator mechanism to perceive fine-grained variations within it. The fusion of these components produces a variable receptive field that is multi-scale and dynamic. A Tide-Ripple Convolutional Neural Network (TR-CNN) is developed to validate this design. To mitigate label noise in training datasets, a total loss function is introduced by combining a NoneTarget with Dynamic Normalization (NTDN) loss and a weighted Sub-center AAM Loss variant, improving model robustness and performance.Results and DiscussionsThe TR-CNN is evaluated on the VoxCeleb1-O/E/H benchmarks. The results show that TR-CNN achieves a competitive balance of accuracy, computation, and parameter efficiency (Table 1). Compared with the strong ECAPA-TDNN baseline, the TR-CNN (C=512, n=1) model attains relative EER reductions of 4.95%, 4.03%, and 6.03%, and MinDCF reductions of 31.55%, 17.14%, and 17.42% across the three test sets, while requiring 32.7% fewer parameters and 23.5% less computation (Table 2). The optimal TR-CNN (C=1 024, n=1) model further improves performance, achieving EERs of 0.85%, 1.10%, and 2.05%. Robustness is strengthened by the proposed total loss function, which yields consistent improvements in EER and MinDCF during fine-tuning (Table 3). Additional evaluations, including ablation studies (Tables 5 and 6), component analyses (Fig. 3 and Table 4), and t-SNE visualizations (Fig. 4), confirm the effectiveness and robustness of each module in the TR-CNN architecture.ConclusionsThis research proposes a simple and effective TR-Conv layer built on the T-RRF mechanism. Experimental results show that TR-Conv forms a more expressive and effective receptive field, reducing parameter count and computational cost while exceeding conventional one-dimensional convolution in speech feature modeling. It also exhibits strong lightweight characteristics and scalability. Furthermore, a total loss function combining the NTDN loss and a Sub-center AAM loss variant is proposed to enhance the discriminability and robustness of speaker embeddings, particularly under label noise. TR-Conv shows potential as a general-purpose module for integration into deeper and more complex network architectures.