桂林电子科技大学学报2026，Vol.46Issue(2)：163-170,8.DOI:10.16725/j.1673-808X.202467
基于特征融合与时延神经网络的说话人确认系统
Speaker verification system based on feature fusion and time-delay neural networks
摘要
Abstract
In response to the problem that single-feature parameters in speaker recognition often fail to fully capture the unique infor-mation of the speaker and the decline in recognition rates in traditional voiceprint recognition systems under complex noisy environ-ments,a speaker confirmation system combining feature fusion and time-delay neural networks has been proposed.The system first analyzes MFCC(mel-frequency cepstral coefficient)and GFCC(gammatone frequency cepstral coefficients)features.MFCC fea-tures are widely used due to their effective representation of the Mel-frequency characteristics of speech signals,while GFCC fea-tures demonstrate better robustness in noisy environments as they mimic the auditory properties of the human cochlea.Subsequently,these two types of feature parameters and their dynamic characteristics are combined to form a high-dimensional feature set.To im-prove the system's processing efficiency,Principal Component Analysis(PCA)technology is used to reduce the dimensionality of the combined feature parameter set.Then,the k-means clustering algorithm is applied to further group these features,thereby con-structing a new type of hybrid feature parameter.Finally,this hybrid feature parameter is applied to the ECAPA-TDNN,a time-de-lay neural network with a channel attention mechanism,for both training and testing.Experimental results show that the proposed method has improved the recognition rate by 27.21%in the specified-5 dB noisy environment compared to traditional single fea-tures,demonstrating better recognition performance.关键词
说话人确认/特征融合/MFCC/GFCC/时延神经网络Key words
speaker verification/feature fusion/MFCC/GFCC/time-delay neural network分类
信息技术与安全科学引用本文复制引用
何青怡,曾庆宁,赵学军..基于特征融合与时延神经网络的说话人确认系统[J].桂林电子科技大学学报,2026,46(2):163-170,8.基金项目
国家自然科学基金(61961009) （61961009）
桂林电子科技大学研究生教育创新计划(2022YCXS042) （2022YCXS042）
