Paper deep dive
MECT: Mixture of Experts with CNN-Transformer Network for Speaker verification
Yu Zheng, Jinghan Peng, ChangHao Zhang, Jian Liu, Weiqiang Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 9/27/2026, 3:40:59 AM
Summary
The paper introduces MECT, a fully supervised speaker verification model that integrates Mixture-of-Experts (MoE) into a CNN-Transformer backbone. It explores four MoE variants (utterance/frame-level, dense/sparse routing) and achieves state-of-the-art performance on VoxCeleb1 and CN-Celeb datasets with low computational complexity. Additionally, it establishes a streaming inference paradigm via causal retraining, maintaining high accuracy at 100ms chunk sizes.
Entities (10)
Relation Signals (6)
MECT → usestechnique → Mixture-of-Experts
confidence 98% · MECT, a speaker verification model that integrates the Mixture-of-Experts (MoE) mechanism
MECT-B2 → achievesstateofthearton → VoxCeleb1
confidence 95% · MECT-B2 achieves state-of-the-art performance on VoxCeleb1
MECT → hasarchitecture → CNN-Transformer
confidence 95% · integrates the Mixture-of-Experts (MoE) mechanism into a CNN-Transformer backbone
MECT → supportsstreaminginference → Causal Retraining
confidence 92% · establish a streaming inference paradigm through causal retraining
MECT → isbasedon → ReDimNet
confidence 90% · MECT adopts a CNN-Transformer architecture based on the Reshape Dimension strategy of ReDimNet
MECT → useslossfunction → SphereFace2
confidence 90% · SphereFace2 loss [7] is used for model training
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In this paper, we propose MECT, a speaker verification model that integrates the Mixture-of-Experts (MoE) mechanism into a CNN-Transformer backbone with optimized block structure and stacking scheme. Specifically, we investigated four MoE variants that span utterance-level and frame-level granularity with dense and sparse routing strategies. The MoE mechanism proves to be effective over the baseline without MoE with only a small increase in parameters. We further scale MECT to a series of model sizes, all maintaining compact parameters and low computational complexity. In particular, MECT-B2 achieves state-of-the-art performance on VoxCeleb1 and delivers strong results on CN-Celeb, demonstrating its effectiveness across diverse datasets. In addition, we establish a streaming inference paradigm through causal retraining, which maintains strong performance at a chunk size of 100ms.
Tags
Links
- Source: https://arxiv.org/abs/2609.24061v1
- Canonical: https://arxiv.org/abs/2609.24061v1
Trouble viewing inline? Open PDF directly →
Full Text
27,946 characters extracted from source content.
Expand or collapse full text
MECT: Mixture of Experts with CNN-Transformer Network for Speaker verification Yu Zheng Jinghan Peng ChangHao Zhang Jian Liu† †thanks: †Corresponding author: rex.lj@antgroup.com https://github.com/ant-research/AntSpeaker Weiqiang Wang Abstract In this paper, we propose MECT, a speaker verification model that integrates the Mixture-of-Experts (MoE) mechanism into a CNN-Transformer backbone with optimized block structure and stacking scheme. Specifically, we investigated four MoE variants that span utterance-level and frame-level granularity with dense and sparse routing strategies. The MoE mechanism proves to be effective over the baseline without MoE with only a small increase in parameters. We further scale MECT to a series of model sizes, all maintaining compact parameters and low computational complexity. In particular, MECT-B2 achieves state-of-the-art performance on VoxCeleb1 and delivers strong results on CN-Celeb, demonstrating its effectiveness across diverse datasets. In addition, we establish a streaming inference paradigm through causal retraining, which maintains strong performance at a chunk size of 100ms. Index Terms: speaker verification, MoE, streaming inference †address: Machine Intelligence, Ant Group, Shanghai, China 1 Introduction Speaker verification (SV) aims to determine whether two speech segments are from the same speaker. In recent years, deep neural network architectures for SV have evolved rapidly. Representative models include 1D convolution TDNN-based frameworks such as ECAPA-TDNN [5, 34] and Next-TDNN [9], CNN-based architectures such as ResNet [8, 32] and its variants [35, 1] including DF-ResNet and Gemini DF-ResNet [17, 18], and CNN–TDNN hybrid structures such as CAM++ [28] and ECAPA2 [25]. Transformer-based networks have also been investigated [33, 29, 3, 30]. ReDimNet [30] introduces a reshape-dimension strategy to effectively integrate CNN, TDNN, and Transformer modules, while ReDimNet2 [31] further improves performance through time-pooled dimension reshaping. Self-supervised pre-trained models (PTMs) have also been applied to speaker verification, with w2v-BERT 2.0-based SV [14] achieving particularly strong performance through fine-tuning. Mixture-of-Experts (MoE) [22] has been widely adopted in large-scale Transformer-based systems, improving scalability through dynamic expert routing. In speaker verification, MoE has only been applied to fine-tuning PTMs [13, 15], where it adaptively combines representations from different layers of the pretrained model. However, MoE has not yet been explored in fully supervised SV. We propose MECT, a fully supervised speaker verification model that introduces the MoE mechanism. Four MoE structures are designed with different expert granularities and routing strategies: utterance-level Dense MoE, utterance-level Sparse MoE, frame-level Dense MoE, and frame-level Sparse MoE, each replacing the first linear projection in the Transformer feed-forward block. MECT adopts a CNN-Transformer architecture based on the Reshape Dimension strategy of ReDimNet, with a redesigned stacking scheme and network structure for better performance. The main contributions of this paper are summarized as follows: 1) We propose MECT, the first MoE-based fully supervised speaker verification model, achieving state-of-the-art performance on Vox1-O, Vox1-E, and Vox1-H with minDCF of 0.012, 0.026, and 0.048, while maintaining low model complexity. 2) We investigate four MoE structures and demonstrate performance gains over the baseline across datasets, with frame-level dense MoE excelling on VoxCeleb and sparse Top-K MoE on CN-Celeb. 3) We establish a streaming inference paradigm for SV through causal retraining, achieving competitive performance at a chunk size of 100ms. Figure 1: The architecture of MECT Block. Left shows the overall block, middle the four MoE variants, and right the CNN and PosEncConv modules. 2 METHODS This section describes the architecture of MECT. We first introduce the basic building block, then detail the MoE layer as the core component and a key contribution of this work. Next, we present the stacking scheme for constructing MECT models at different scales to accommodate diverse deployment requirements. Finally, we describe the streaming speaker embedding extraction scheme. 2.1 Basic Block of MECT As illustrated in Figure1, a MECT block consists of a CNN and a Transformer equipped with MoE. The CNN and Transformer are connected via the reshape and time-pooled dimension-reshaping methods adopted from ReDimNet and ReDimNet2[30, 31]. Beyond MoE, our block design also differs from ReDimNet in several aspects that contribute to further performance gains. Residual connection. We adopt ResNet blocks for the CNN, but place the shortcut addition after BatchNorm and ReLU rather than before, as shown in the CNNBlock module of Figure1. This design consistently yields better performance in our experiments. Position encoding. We employ a two-layer 1D convolution as the position encoding, as illustrated in the PosEncConv of Figure1. Downsampling and channel expansion. Both temporal downsampling and channel expansion are handled by convolutions within the ResNet block, rather than by a separate convolution layer before it as in ReDimNet. This design simplifies the module structure and leads to more stable training. Block connection. We connect blocks via a simple shortcut rather than the dense connections used in ReDimNet. In our experiments, dense connections did not improve performance while increasing computational overhead. 2.2 Mixture-of-Experts layer We explore four MoE variants that differ in aggregation granularity (utterance-level vs. frame-level) and routing strategy (Dense vs. Sparse), as shown in Figure 1. Let ∈ℝT×dx ^T× d denote the input feature sequence to the MoE layer, where T and d are the temporal length and feature dimension, respectively. Let t∈ℝdx_t ^d denotes a single frame. 2.2.1 Frame-level MoE In the Dense MoE, the gating network is a linear layer parameterized by g∈ℝd×NW_g ^d× N, where N denotes the number of experts. Its output tG_t serves as the routing weights for each expert at frame t. Each expert is a linear layer parameterized by eiW_e^i, and f(⋅)f(·) denotes the activation function. The output of the MoE layer, ty_t, is computed as t=Softmax(tg),t=∑i=1Nti⋅f(tei).G_t=Softmax(x_tW_g), _t= _i=1^NG_t^i· f(x_tW_e^i). (1) In the Sparse MoE, only the Top-K experts are activated for routing. tS_t and th_t denote the indices and the corresponding logits of the selected experts, respectively. It is formulated as follows: t,t=TopK(tg,K),S_t,h_t=TopK(x_tW_g,K), (2) t=Softmax(t),t=∑i∈tti⋅f(tei).G_t=Softmax(h_t), _t= _i _tG_t^i· f(x_tW_e^i). (3) 2.2.2 Utterance-level MoE In the utterance-level variants, the gating decision is made based on the mean-pooled utterance representation ¯=1T∑t=1Tt x= 1T _t=1^Tx_t, producing a single set of routing weights shared across all frames. The utterance-level Dense MoE is formulated as follows: =Softmax(¯g),t=∑i=1Ni⋅f(tei).G=Softmax( xW_g), _t= _i=1^NG^i· f(x_tW_e^i). (4) And the utterance-level Sparse MoE is formulated as follow: ,=TopK(¯g,K),S,g=TopK( xW_g,K), (5) =Softmax(),t=∑i∈i⋅f(tei).G=Softmax(g), _t= _i G^i· f(x_tW_e^i). (6) 2.3 MECT architecture Table 1: Architecture of MECT. Layer name Structure Output Input - (1, 80, T ) Head Conv2D, stride 1 (C, 80, T) Stage1 Block × 2, stride 1 (C, 80, T) Stage2 Block × 4, stride (2,1),(1,2) × 3 (2 × C, 40, T) Stage3 Block × 5, stride (2,1),(1,2) × 4 (4 × C, 20, T) Stage4 Block × 3, stride (2,1),(1,2) × 2 (8 × C, 10, T) Fusion Weight Dense (80 × C, T) Pooling ASTP 80 × C Projection Linear 192 Loss SphereFace2 Num. Speakers Table 1 details the MECT architecture. Each stage stacks multiple MECT blocks, where the first block applies frequency downsampling, and the remaining blocks apply temporal downsampling, both via convolutional stride. Outputs of all stages are fused via a weighted dense layer before pooling. We provide four MECT model sizes controlled by two parameters: the input channel size C and the number of ResBlocks per block M, with a fixed Transformer hidden dimension d=64d=64. A1 and B1 use M=2M=2, while A2 and B2 use M=3M=3; A1 and A2 use C=16C=16, while B1 and B2 use C=32C=32. We use 80-dimensional mean-normalized log Mel filter-bank features extracted with a 25 ms window and 10 ms frame shift over 20–7600 Hz at a sampling rate of 16 kHz. Attentive Statistics Pooling (ASTP) [20] is employed for utterance-level aggregation, and SphereFace2 loss [7] is used for model training. 2.4 Streaming speaker embedding extract We compare three streaming strategies, denoted M1, M2, and M3. M1 (embedding averaging) simply averages the embedding of the current chunk with those of all prior chunks. M2 (feature concatenation) concatenates all accumulated features before pooling, then passes them through the subsequent network to produce the embedding. M3 (causal retraining) requires retraining the model under a causal architecture, as described below. We first convert the network to a causal architecture. For convolution layers, padding is applied only on the left side of the temporal dimension, and for multi-head attention (MHA), a causal mask restricts each frame to attend only to preceding frames. The model is retrained accordingly. During streaming inference (Figure 2), no mean normalization is applied to the input fbank features. Conv2d layers cache 2 frames (or 6 for temporal stride 2) for receptive field alignment, while Conv1d layers cache 112 frames to cover the accumulated stride, both concatenated with the current input along the temporal dimension. For MHA, the KV states of all previous steps are concatenated with the current step before attention. For ASTP, the input frame features and intermediate attention logits (feature before softmax) from prior chunks are cached and concatenated with the current chunk to compute global mean and standard deviation. Since only the KV cache and ASTP rely on full history and both are lightweight, the additional cost remains modest. In practice, both caches can be capped to bound memory usage. Our experiments do not apply this cap to focus on validating the streaming algorithm. 3 Experiment 3.1 Datasets We evaluated MECT on multiple datasets. For VoxCeleb model training, we used the VoxCeleb2 [4] development set, either alone or combined with VoxBlink2 [16]. SV performance was evaluated with VoxCeleb1-Cleaned protocols: Vox1-O, Vox1-E, Vox1-H and Vox21-val. Additionally, the development sets of CN-Celeb1 and CN-Celeb2 [6] were used for training, with evaluation on CN-Celeb Test. These experiments span different languages, scales, and data sources, validating the effectiveness of the proposed method. Figure 2: Streaming inference pipeline of MECT. Illustrating the causal processing and multi-level caching across consecutive chunks. Table 2: Evaluation results on different MoE types. Type No.E Vox1-O Vox1-E Vox1-H Vox21-val Avg EER DCF No MoE – 0.33 0.48 0.85 1.68 0.84 0.064 F-Dense 2 0.34 0.49 0.89 1.71 0.86 0.066 F-Dense 3 0.27 0.49 0.89 1.65 0.83 0.061 F-Dense 4 0.27 0.46 0.85 1.60 0.80 0.059 F-Dense 5 0.35 0.51 0.88 1.61 0.84 0.061 U-Dense 3 0.30 0.47 0.87 1.62 0.82 0.062 U-Dense 4 0.36 0.48 0.89 1.60 0.83 0.067 F-Top1 4 0.36 0.48 0.88 1.61 0.83 0.064 U-Top1 4 0.37 0.48 0.87 1.61 0.83 0.061 F-Top4 6 0.33 0.49 0.89 1.57 0.82 0.066 F-Top4 8 0.30 0.48 0.87 1.55 0.80 0.063 F-Top4 16 0.31 0.47 0.86 1.57 0.80 0.061 F-Top4 32 0.33 0.48 0.88 1.51 0.80 0.063 Table 3: Evaluation results on VoxCeleb1-Cleaned. * indicates results obtained using the VoxCeleb2 and VoxBlink2 datasets for training. GMACs were measured on 2-s segments counting only the backbone, with ReDimNet2 using the official open-source model. Model Params GMACs LMF AS-Norm Vox1-O Vox1-E Vox1-H EER(%) minDCF EER(%) minDCF EER(%) minDCF ECAPA (C=512) [5, 17] 6.2M 1.04 × ✓ 0.94 0.092 1.21 0.129 2.20 0.205 CAM++ [28, 27] 7.2M 1.15 ✓ × 0.71 0.109 0.85 0.095 1.66 0.165 ReDimNet2-B3 [31] 2.58M 2.58 ✓ × 0.42 0.038 0.66 0.068 1.22 0.120 MECT-A1 3.78M 3.02 ✓ × 0.44 0.038 0.60 0.060 1.05 0.108 Gemini DF-ResNet114 [18] 6.5M 5.0 × ✓ 0.69 0.067 0.86 0.097 1.49 0.144 ReDimNet2-B4 [31] 4.52M 4.48 ✓ × 0.37 0.040 0.58 0.060 1.07 0.102 MECT-A2 4.12M 4.11 ✓ × 0.37 0.032 0.54 0.056 0.97 0.096 ResNet293 [27] 28.6M 28.10 ✓ × 0.53 0.057 0.71 0.072 1.30 0.127 ECAPA2 [25] 27.1M 187.0 ✓ × 0.44 0.041 0.62 0.066 1.15 0.114 WavLM Large [10] 319.3M 28.5 ✓ ✓ 0.37 0.059 0.50 0.055 1.01 0.099 MECT-B1 8.26M 10.35 ✓ × 0.37 0.029 0.50 0.051 0.94 0.092 ReDimNet2-B6 [31] 10.45M 13.11 ✓ × 0.29 0.027 0.52 0.050 0.99 0.104 MECT-B2 9.57M 14.43 ✓ × 0.27 0.024 0.46 0.048 0.85 0.082 W2V-BERT 2.0* [14] 587M 57.90 ✓ × 0.14 0.020 0.31 0.032 0.73 0.071 +AS-Norm & QMF ✓ ✓ 0.12 0.025 0.27 0.028 0.55 0.051 MECT-B2* 9.57M 14.43 ✓ × 0.23 0.013 0.29 0.028 0.54 0.052 +AS-Norm & QMF ✓ ✓ 0.22 0.012 0.28 0.026 0.52 0.048 3.2 Training Experiments were conducted under three data settings. All models were trained in two stages: full training followed by large-margin fine-tuning (LMF) [26]. All experiments used the same hardware and PyTorch environment, and random seeds were fixed across ablation studies to isolate the effect of each component. VoxCeleb2. VoxCeleb2 was augmented with two-fold speed perturbation (0.9 and 1.1) [11], yielding 17,982 speakers. Two-second segments were randomly extracted with noise and reverberation augmentation via MUSAN [23] and RIR [24] following [27]. SGD was used with momentum 0.9, weight decay 2e-5, and batch size 512. The learning rate warmed up to 0.6 over 4 epochs, then decayed to 4e-4 at epoch 120. The SphereFace2 margin was set to 0.0 for 20 epochs, raised to 0.2 over the next 20, and held constant. In the LMF stage, speed perturbation was removed (5,994 speakers), segment length increased to 6 s, the margin was fixed at 0.3, and the learning rate was reduced to 2e-4 for 7 epochs. CN-Celeb. The same training recipe was adopted with a maximum learning rate of 1.6. The model from epoch 50 of the first stage was used for LMF, with a learning rate of 8e-4. VoxCeleb2 & VoxBlink2. Speed perturbation was not applied. The first-stage training followed the same settings as the VoxCeleb2 experiment, while the second stage used VoxCeleb2 alone for LMF. 3.3 Evaluation The performance was evaluated using Equal Error Rate (EER) and minimum Decision Cost Function (minDCF) with CFA=CMiss=1C_FA=C_Miss=1, Ptarget=0.01P_target=0.01, except on Vox21-val Ptarget=0.05P_target=0.05. AS-Norm [19] and QMF [26] were applied for selected systems using VoxCeleb2 development set as backend. 3.4 Streaming inference test During streaming evaluation, the input audio was processed chunk by chunk, and the final embedding was obtained after the last chunk. For M1 and M2, adjacent chunks overlap by half the chunk size for better performance. 4 Results Table 2 reports representative results from our full ablation study of different MoE types and expert counts, all trained on VoxCeleb2 and evaluated on the VoxCeleb1 test sets. “No MoE” denotes the baseline without MoE modules. Among all configurations, Frame-level Dense MoE with 4 experts achieves the best results, outperforming the baseline across all test sets with relative gains of 4.7% and 7.8% on average EER and minDCF. Notably, this improvement comes at a negligible parameter cost: the baseline contains 9.40M parameters versus 9.57M for the best MoE configuration. Figure 3: Statistics of expert weights across network blocks and phonemes. Table 3 compares MECT with publicly available models on the VoxCeleb1 protocols, where MECT uses frame-level Dense MoE with 4 experts. In group 2, MECT-A2 outperforms ReDimNet2-B4 across all three test sets with the fewest parameters and MACs, achieving particularly notable gains on Vox1-H (9.4% and 5.9% relative improvements on EER and minDCF). In group 3, MECT-B2 surpasses ReDimNet2-B6 with fewer parameters and comparable MACs, delivering consistent improvements on all sets and a 14.1% relative gain on Vox1-H minDCF, while the smaller MECT-B1 already edges past ReDimNet2-B6. With VoxBlink2 training, MECT-B2 outperforms w2v-bert2.0 using far fewer resources, achieving substantial improvements on Vox1-O, Vox1-E, and Vox1-H with relative gains of 10.3% and 24.4% on average EER and minDCF. After AS-Norm and QMF, MECT-B2 further improves and reaches state-of-the-art performance, with minDCFs of 0.012, 0.026, and 0.048. Table 4 reports results on CN-Celeb with models trained only on CN-Celeb. MECT-B2 uses frame-level sparse Top-4 MoE with 8 experts and achieves relative improvements of 6.7% on EER and 7.5% on minDCF over ReDimNet2-B6 trained under the same settings. Figure3 shows the MoE routing behavior in MECT-B2 on VoxCeleb1. Using the Allosaurus[12] phoneme recognizer, we correlate routing weights with phoneme categories in the first block. As shown in Figure3(b), E2 assigns high weights (0.65–0.88) to fricatives and affricates[21] (e.g., /s, z, , f, , t, d/), whose high-frequency noise distinguishes them from other phonemes, while vowels (e.g., /e, o, u, æ/) are routed mainly to E0 with moderate weights (0.40–0.55). This specialization emerges without phoneme supervision, suggesting the MoE mechanism naturally captures the acoustic structure of speech. As shown in Table 5, at a chunk size of 1.0s, M2 achieves the best streaming performance. However, at 0.1s, M1 and M2 degrade severely to 12.34% and 7.50% on Vox1-O, while M3 remains at 0.38%, 0.60%, and 1.06% on Vox1-O, Vox1-E, and Vox1-H, closely matching its offline baseline. This demonstrates the effectiveness of causal retraining for short-chunk streaming. Table 4: Evaluation results on CN-Celeb. Model Params GMACs LMF AS-Norm CN-Celeb Test EER(%) DCF CAM++ [2] 7.2M 1.15 ✓ × 6.30 0.370 ERes2NetV2 [2] 17.8M 12.60 ✓ × 6.04 0.362 ResNet221 [27] 23.9M 21.29 ✓ ✓ 5.66 0.330 ReDimNet2-B6 12.3M 13.05 ✓ × 5.22 0.321 MECT-B2 9.81M 14.46 ✓ × 4.87 0.297 w/o MoE 9.57M 14.43 ✓ × 5.01 0.303 Table 5: Streaming inference results on VoxCeleb1. MECT-B2† denotes the causally retrained model. Model Method Chunk Vox1-O Vox1-E Vox1-H MECT-B2 Offline – 0.27 0.46 0.85 Stream M1 0.1s 12.34 12.36 17.40 0.2s 4.50 4.45 6.87 1.0s 0.47 0.67 1.15 Stream M2 0.1s 7.50 7.55 10.91 0.2s 2.04 2.14 3.50 1.0s 0.39 0.58 1.03 MECT-B2† Offline – 0.38 0.59 1.05 Stream M3 0.1s 0.38 0.60 1.06 5 Conclusions We proposed MECT, a MoE-based speaker verification model. Four MoE structures were investigated, and performance gains over the baseline were observed across datasets. MECT achieves state-of-the-art performance on VoxCeleb1. A causal retraining strategy enables streaming inference at 0.1 s chunks with near-offline accuracy, substantially outperforming non-causal streaming methods. Visualization of expert routing indicates that MoE learns phoneme-level specialization to some extent, without explicit supervision. References [1] Y. Chen, S. Zheng, H. Wang, L. Cheng, Q. Chen, and J. Qi (2023) An Enhanced Res2Net with Local and Global Feature Fusion for Speaker Verification. In Proc. Interspeech, p. 2228–2232. External Links: Document, ISSN 2958-1796 Cited by: §1. [2] Y. Chen, S. Zheng, H. Wang, L. Cheng, T. Zhu, R. Huang, C. Deng, Q. Chen, S. Zhang, W. Wang, and X. Li (2025) 3D-speaker-toolkit: an open-source toolkit for multimodal speaker verification and diarization. In Proc. ICASSP, Vol. , p. 1–5. External Links: Document Cited by: Table 4, Table 4. [3] J. Choi, J. Yang, and J. Chang (2024) Efficient lightweight speaker verification with broadcasting cnn-transformer and knowledge distillation training of self-attention maps. IEEE/ACM 32 (), p. 4580–4595. External Links: Document Cited by: §1. [4] J. S. Chung, A. Nagrani, and A. Zisserman (2018) VoxCeleb2: Deep Speaker Recognition. In Interspeech 2018, p. 1086–1090. External Links: Document, ISSN 2958-1796 Cited by: §3.1. [5] B. Desplanques, J. Thienpondt, and K. Demuynck (2020) ECAPA-tdnn: emphasized channel attention, propagation and aggregation in tdnn based speaker verification. In Proc. Interspeech, p. 3830–3834. External Links: Document Cited by: §1, Table 3. [6] Y. Fan, J.W. Kang, L.T. Li, K.C. Li, H.L. Chen, S.T. Cheng, P.Y. Zhang, Z.Y. Zhou, Y.Q. Cai, and D. Wang (2020) CN-celeb: a challenging chinese speaker recognition dataset. In Proc. ICASSP, Vol. , p. 7604–7608. Cited by: §3.1. [7] B. Han, Z. Chen, and Y. Qian (2023) Exploring binary classification loss for speaker verification. In Proc. ICASSP, Vol. , p. 1–5. External Links: Document Cited by: §2.3. [8] K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 770–778. Cited by: §1. [9] H. Heo, U. Shin, R. Lee, Y. Cheon, and H. Park (2024) NeXt-tdnn: modernizing multi-scale temporal convolution backbone for speaker verification. In Proc. ICASSP, Vol. , p. 11186–11190. External Links: Document, ISSN 2379-190X Cited by: §1. [10] J. S. Kim, H. J. Park, W. Shin, and S. W. Han (2025) Rethinking Leveraging Pre-Trained Multi-Layer Representations for Speaker Verification. In Interspeech 2025, p. 3713–3717. External Links: Document, ISSN 2958-1796 Cited by: Table 3. [11] T. Ko, V. Peddinti, D. Povey, and S. Khudanpur (2015) Audio augmentation for speech recognition. In Proc. Interspeech, p. 3586–3589. External Links: Document, ISSN 2958-1796 Cited by: §3.2. [12] X. Li, F. Metze, D. R. Mortensen, A. W. Black, and W. Xiong (2020) Allosaurus: a multilingual pretrained universal phone recognizer. In Proc. Interspeech, Cited by: §4. [13] Y. Li, Y. Yu, Y. Tu, S. Deng, and W. Gan (2026) Enhancing speaker verification with layer-wise mixture-of-experts on pre-trained models. In Proc. ICASSP, Vol. , p. 19032–19036. External Links: Document Cited by: §1. [14] Z. Li, M. Cheng, and M. Li (2026) Enhancing speaker verification with w2v-bert 2.0 and knowledge distillation guided structured pruning. In Proc. ICASSP, Vol. , p. 16462–16466. External Links: Document Cited by: §1, Table 3. [15] Z. Li, M. Mak, M. Pilanci, H. Lee, C. Gan, J. Sheng, and H. Meng (2026) Toward a unified perspective on parameter-efficient fine tuning for speaker verification. IEEE Transactions on Audio, Speech and Language Processing 34 (), p. 2276–2289. External Links: Document Cited by: §1. [16] Y. Lin, M. Cheng, F. Zhang, Y. Gao, S. Zhang, and M. Li (2024) VoxBlink2: A 100K+ Speaker Recognition Corpus and the Open-Set Speaker-Identification Benchmark. In Interspeech 2024, p. 4263–4267. External Links: ISSN 2958-1796 Cited by: §3.1. [17] B. Liu, Z. Chen, S. Wang, H. Wang, B. Han, and Y. Qian (2022) DF-ResNet: Boosting Speaker Verification Performance with Depth-First Design. In Proc. Interspeech, p. 296–300. External Links: Document, ISSN 2958-1796 Cited by: §1, Table 3. [18] T. Liu, K. A. Lee, Q. Wang, and H. Li (2024) Golden gemini is all you need: finding the sweet spots for speaker verification. IEEE/ACM 32 (), p. 2324–2337. External Links: Document Cited by: §1, Table 3. [19] P. Matějka, O. Novotný, O. Plchot, L. Burget, M. D. Sánchez, and J. Černocký (2017) Analysis of Score Normalization in Multilingual Speaker Recognition. In Interspeech 2017, p. 1567–1571. External Links: Document, ISSN 2958-1796 Cited by: §3.3. [20] K. Okabe, T. Koshinaka, and K. Shinoda (2018) Attentive Statistics Pooling for Deep Speaker Embedding. In Proc. Interspeech, p. 2252–2256. External Links: Document, ISSN 2958-1796 Cited by: §2.3. [21] J. M. Pickett and S. R. Morris (2000) The acoustics of speech communication: fundamentals, speech perception theory, and technology. The Journal of the Acoustical Society of America 108 (4), p. 1373–1374. External Links: Document Cited by: §4. [22] N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. V. Le, G. E. Hinton, and J. Dean (2017) Outrageously large neural networks: the sparsely-gated mixture-of-experts layer. ArXiv abs/1701.06538. External Links: Link Cited by: §1. [23] D. Snyder, G. Chen, and D. Povey (2015) MUSAN: a music, speech, and noise corpus. arXiv preprint arXiv:1510.08484. Cited by: §3.2. [24] I. Szöke, M. Skácel, L. Mošner, J. Paliesek, and J. Černocký (2019) Building and evaluation of a real room impulse response dataset. IEEE Journal of Selected Topics in Signal Processing 13 (4), p. 863–876. External Links: Document Cited by: §3.2. [25] J. Thienpondt and K. Demuynck (2023) ECAPA2: a hybrid neural network architecture and training strategy for robust speaker embeddings. In 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), Vol. , p. 1–8. External Links: Document Cited by: §1, Table 3. [26] J. Thienpondt, B. Desplanques, and K. Demuynck (2021) The idlab voxsrc-20 submission: large margin fine-tuning and quality-aware score calibration in dnn based speaker verification. In Proc. ICASSP, p. 5814–5818. Cited by: §3.2, §3.3. [27] H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xiang, Y. Deng, and Y. Qian (2023) Wespeaker: a research and production oriented speaker embedding learning toolkit. In Proc. ICASSP, Vol. , p. 1–5. External Links: Document Cited by: §3.2, Table 3, Table 3, Table 4. [28] H. Wang, S. Zheng, Y. Chen, L. Cheng, and Q. Chen (2023) CAM++: A Fast and Efficient Network for Speaker Verification Using Context-Aware Masking. In Proc. Interspeech, p. 5301–5305. External Links: Document, ISSN 2958-1796 Cited by: §1, Table 3. [29] R. Wang, J. Ao, L. Zhou, S. Liu, Z. Wei, T. Ko, Q. Li, and Y. Zhang (2022) Multi-view self-attention based transformer for speaker recognition. In Proc. ICASSP, Vol. , p. 6732–6736. External Links: Document Cited by: §1. [30] I. Yakovlev, R. Makarov, A. Balykin, P. Malov, A. Okhotnikov, and N. Torgashov (2024) Reshape Dimensions Network for Speaker Recognition. In Proc. Interspeech, p. 3235–3239. External Links: Document, ISSN 2958-1796 Cited by: §1, §2.1. [31] I. Yakovlev and A. Okhotnikov (2026) ReDimNet2: scaling speaker verification via time-pooled dimension reshaping. arXiv preprint arXiv:2603.11841. External Links: Link Cited by: §1, §2.1, Table 3, Table 3, Table 3. [32] H. Zeinali, S. Wang, A. Silnova, P. Matějka, and O. Plchot (2019) But system description to voxceleb speaker recognition challenge 2019. arXiv preprint arXiv:1910.12592. Cited by: §1. [33] Y. Zhang, Z. Lv, H. Wu, S. Zhang, P. Hu, Z. Wu, H. Lee, and H. Meng (2022) MFA-conformer: multi-scale feature aggregation conformer for automatic speaker verification. In Proc. Interspeech, p. 306–310. Cited by: §1. [34] Z. Zhao, Z. Li, W. Wang, and P. Zhang (2023) PCF: ecapa-tdnn with progressive channel fusion for speaker verification. In Proc. ICASSP, p. 1–5. Cited by: §1. [35] T. Zhou, Y. Zhao, and J. Wu (2021) ResNeXt and res2net structures for speaker verification. In 2021 IEEE Spoken Language Technology Workshop (SLT), Vol. , p. 301–307. External Links: Document Cited by: §1.