1 School of Information and Communication, Guilin University of Electronic Technology, Guilin 541004, China
2 Key Laboratory of Cognitive Radio and Information Processing, Ministry of Education (Guilin University of Electronic Technology), Guilin 541004, China
In recent years, deep learning has made significant progress in speaker verification tasks. However, existing models often suffer from large parameter sizes, limited temporal modeling capability, and poor deployability on edge devices. To address these issues, a lightweight speaker verification model incorporating grouped modeling and multi-dimensional attention mechanisms was proposed. The model was built upon an improved time-delay neural network (TDNN) architecture and introduced a channel-split temporal encoding block (CSTE-Block) and a grouped scaled Res2Net block (GSRes2Block) to enhance frame-level temporal modeling and multi-scale feature representation. In addition, a speaker embedding refinement module block (SERMBlock) and multi-dimensional attention were employed to improve global contextual modeling and feature focusing. A global speaker embedding was generated via attentive statistical pooling and jointly optimized using an additive angular margin Softmax classifier and supervised contrastive loss to enhance the discriminative power of the embedding space. Experiments on the CN-Celeb dataset demonstrate that the proposed model achieves an equal error rate (EER) of 14.82% and a minimum detection cost function (minDCF) of 0.694 8, with a parameter reduction of approximately 69.4% compared to the ECAPA-TDNN(emphasized channel attention, propagation and aggregation in time delay neural network) baseline. The model also exhibits strong cross-lingual robustness on the VoxCeleb1 evaluation set. Complexity analysis further confirms that the model offers a favorable trade-off between performance and computational efficiency, showing excellent potential for real-world deployment.
ZHENG Zhan-heng, LI Jia-qi, WANG Jian, LI Qi, ZENG Qing-ning.
Lightweight Speaker Verification Method Based on Fusion of Grouped Modeling and Multi-dimensional Attention Mechanism[J].
Science Technology and Engineering, 2026, 26(27): 11804-11811 DOI:10.12404/j.issn.1671-1815.2600566
早期以高斯混合模型与通用背景模型构建的i-vector系统[7],通过提取如梅尔频率倒谱系数(mel-frequency cepstral coefficients, MFCC)或对数梅尔滤波器组能量(filter banks, FBank)等声学特征获得低维说话人向量,配合PLDA(probabilistic linear discriminant analysis)判别器完成匹配判定。尽管该方法在部分场景下表现稳定,但在变长语音、跨信道或复杂环境下的泛化能力受限。
随着深度学习的发展,x-vector系统[8]基于时间延迟神经网络[9]引入帧级建模与统计池化机制,显著提升了说话人特征的判别性。后续研究引入ResNet[10]提升特征表达深度,ECAPA-TDNN(emphasized channel attention, propagation and aggregation in time delay neural network) [11]则融合多尺度残差结构、通道注意力机制[12]与多层特征聚合[13],有效增强了嵌入性能,成为当前主流方法之一。
SharmaR, GovindD, MishraJ, et al. Milestones in speaker recognition[J]. Artificial Intelligence Review, 2024, 57(3): 58.
[2]
KatavA, MosheY, CohenI. A framework for robust speaker verification in highly noisy environments leveraging both noisy and enhanced audio[C]// 2025 33rd European Signal Processing Conference (EUSIPCO). New York: IEEE, 2025: 41-45.
[3]
WengS, YangL, MaoJ. Effective modeling of critical contextual information for TDNN-based speaker verification[J]. arXiv Preprint, arXiv: 2509. 09932, 2025.
WangMeixi, WangHuapeng, YanDaoshen, et al. Forensic automatic speaker recognition based on enhanced ECAPA-TDNN[J]. Science Technology and Engineering, 2024, 24(27): 11763-11773.
[10]
SongY, LiuF, ShenT. PLDA in i-vector based underwater acoustic signals classification[J]. Ships and Offshore Structures, 2024, 19(3): 366-374.
[11]
ZhangY, LiuL. Multi-task learning for X-vector based speaker recognition[J]. International Journal of Speech Technology, 2023, 26(4): 817-823.
[12]
PeddintiV, PoveyD, KhudanpurS. A time delay neural network architecture for efficient modeling of long temporal contexts[C]// Interspeech. Dresden: International Speech Communication Association, 2015: 3214-3218.
[13]
ZhouT, ZhaoY, WuJ. Resnext and res2net structures for speaker verification[C]// 2021 IEEE Spoken Language Technology Workshop (SLT). New York: IEEE, 2021: 301-307.
[14]
DesplanquesB, ThienpondtJ, DemuynckK. ECAPA-TDNN: emphasized channel attention, propagation and aggregation in TDNN based speaker verification[J]. arXiv Preprint, arXiv: 2005. 07143, 2020.
[15]
HuJ, ShenL, SunG. Squeeze-and-excitation net-works[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2018: 7132-7141.
[16]
LiuT, DasR K, LeeK A, et al. MFA: TDNN with multi-scale frequency-channel attention for text-independent speaker verification with short utterances[C]// ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). New York: IEEE, 2022: 7517-7521.
[17]
TesfaiH, SalehH, Al-QutayriM, et al. Lightweight shufflenet based on CNN for arrhythmia classification[J]. IEEE Access, 2022, 10: 111842-111854.
[18]
MaN, ZhangX, ZhengH T, et al. Shufflenet v2: practical guidelines for efficient CNN architecture design[C]//Proceedings of the European Conference on Computer Vision (ECCV). Berlin: Springer International Publishing, 2018: 116-131.
[19]
LiZ, LiuY, LiL, et al. Additive phoneme-aware margin softmax loss for language recognition[J]. arXiv Preprint, arXiv: 2106. 12851, 2021.
[20]
KhoslaP, TeterwakP, WangC, et al. Supervised contrastive learning[J]. Advances in Neural Information Processing Systems, 2020, 33: 18661-18673.
[21]
ÅgrenW. The NT-Xent loss upper bound[J]. arXiv Preprint, arXiv: 2205. 03169, 2022.
[22]
FanY, KangJ W, LiL T, et al. CN-Celeb: a challenging Chinese speaker recognition dataset[C]// ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). New York: IEEE, 2020: 7604-7608.
[23]
NagraniA, ChungJ S, XieW, et al. VoxCeleb: large-scale speaker verification in the wild[J]. Computer Speech & Language, 2020, 60: 101027.
[24]
JiahongL, JieB, YingshuangC, et al. An adaptive resnet based speaker recognition in radio communication[C]// 2021 IEEE International Conference on Emergency Science and Information Technology (ICESIT). New York: IEEE, 2021: 161-164.
[25]
ChenY, ZhengS, WangH, et al. An enhanced Res2Net with local and global feature fusion for speaker verification[J]. arXiv Preprint, arXiv: 2305. 12838, 2023.
[26]
WangH, ZhengS, ChenY, et al. Cam++: a fast and efficient network for speaker verification using context-aware masking[J]. arXiv Preprint, arXiv: 2303. 00332, 2023.