基于听觉特性和发声特性的语种识别  

Language identification based on auditory and vocal characteristics

作　　者：华英杰 朵琳[1] 刘晶 邵玉斌[1] HUA Ying-jie;DUO Lin;LIU Jing;SHAO Yu-bin(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,Yunnan,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500

出　　处：《云南大学学报（自然科学版）》2023年第4期807-814,共8页Journal of Yunnan University(Natural Sciences Edition)

基　　金：国家自然科学基金(61962032);云南省科技厅优秀青年项目(202001AW07000).

摘　　要：针对现有的方法在低信噪比环境下语种识别性能不佳,提出了一种耳蜗滤波系数和声道冲激响应频谱参数相互融合的语种识别方法.该方法表征了人的耳蜗听觉特性和发声特性,首先提取模拟人耳听觉特性的耳蜗滤波系数,再融合表征人的发声特性的声道冲激响应频谱参数,最后采用高斯混合通用背景模型对所提方法在语种识别上进行测试.实验结果表明,在4种信噪比环境下,该方法优于其他对比方法;相对于基于深度学习的对数Mel尺度滤波器能量特征,识别正确率提升了16.1%,与其他方法相比有较大程度的提升.Aiming at the poor performance of the existing methods in language identification in the low signal-to-noise ratio environment,a language identification method is proposed,which integrates the cochlear filter coefficients and the spectral parameters of the vocal tract impulse response.This method characterizes human vocalization characteristics and human hearing characteristics.Firstly,the cochlear filter coefficients that simulate the auditory characteristics of the human ear are fused.Then the spectral parameters of the vocal tract impulse response that characterize the characteristics of human vocalization are extracted.Finally,the Gaussian mixture general background model is used to test the proposed method in language identification.The experimental results show that in the four signal-to-noise ratio environments,this method is superior to other comparison methods.Compared with the logarithmic Mel-scale filter energy feature based on deep learning,the identification accuracy is improved by 16.1%,which is also very good compared to other methods.

关 键 词：语种识别 耳蜗滤波系数 声道冲激响应频谱参数 高斯混合通用背景模型 

分 类 号：TN912.3[电子电信—通信与信息系统]