短波信道下基于鲁棒语音特征参数的身份识别方法  

A method for extracting robust speech feature parameters in HF channel

作　　者：王亨佳 翁呈祥[1] 胡乔林[1] 刘康 WANG Hengjia;WENG Chengxiang;HU Qiaolin;LIU Kang(Air Force EarlyWarning Academy,Wuhan 430019, China)

机构地区：[1]空军预警学院

出　　处：《空军预警学院学报》2019年第4期281-286,共6页Journal of Air Force Early Warning Academy

摘　　要：针对传统语音特征参数经短波信道传输后严重变形,导致语音身份识别性能下降的问题,提出一种基于线性加权梅尔倒谱系数(LWMFCC)的鲁棒短波语音身份识别方法.首先根据F比方法研究短波信道对线性预测系数(LPCC)和梅尔倒谱系数(MFCC)的影响情况,从声道频率响应的角度分析这两种特征参数变形的共性;然后结合留数归一法和线性预测方法估计语音功率谱,并将功率谱经过梅尔滤波、离散余弦转换和相对谱滤波后,提取出自适应补偿信道影响的特征LWMFCC;最后采用高斯混合模型实现身份识别.实验结果表明,该方法提取的特征对短波信道的影响具有稳健特性,能显著提高身份识别率.Aiming at the problem that the severe deformation of the traditional speech feature parameters after HF channel transmission results in degradation of speech recognition performance, this paper proposes a robust short-wave speech identification method based on linear weighted Mel-frequency cepstrum coefficient (LWMFCC). Firstly, according to the F-ratio method, the paper makes a study on the influence of short-wave channel on linear prediction cepstrum coefficient (LPCC)) and Mel-frequency cepstrum coefficient (MFCC), and analyzes the commonness of these two kinds of parameter deformation from the perspective of channel frequency response. Then the paper combines the residue normalization method and linear prediction technique to estimate the power spectrum of speech, and extracts the characteristic LWMFCC of the adaptive compensation channel after the power spectrum goes through Mel filter, discrete cosine transformation and the relative spectral filtering. Finally, the Gaussian mixture model is used to realize the identification. Experimental results show that the features extracted by this method are robust to the influence of HF channel, and can significantly improve the identification rate.

关 键 词：短波信道 语音特征参数 留数归一化 RASTA滤波 

分 类 号：TN912.3[电子电信—通信与信息系统]