基于HHT瞬时能频值的耳语音端点检测  

Endpoint detection based on instantaneous energy frequency value of HHT for whispered speech

作　　者：谈雪丹[1] 顾济华[1] 赵鹤鸣[2] 陶智[1] 韩韬[1] 吴俊[1] 

机构地区：[1]苏州大学物理科学与技术学院,江苏苏州215006 [2]苏州大学电子信息学院,江苏苏州215021

出　　处：《计算机工程与应用》2010年第29期147-150,188,共5页Computer Engineering and Applications

基　　金：国家自然科学基金(No.60572076);江苏省高校自然科学研究计划基金资助项目(No.05JKB510113)~~

摘　　要：由于耳语音信噪比较低,采用传统的算法进行耳语音端点检测存在正确率低、抗噪性能差等问题。提出了一种基于希尔伯特-黄变换瞬时能频值的耳语音端点检测的算法。运用希尔伯特-黄变换,分离出耳语音的瞬时幅值与频率,提取基于时间-能量-频率的特征参数瞬时能频值,利用该特征值对耳语音和噪声进行区分,进行端点检测。对700个信噪比为2～10dB的耳语音测试样本进行仿真实验,该算法检测的起点正确率与终点正确率均高于零能积法、熵法以及拟和特征法。实验表明,该算法适应于多种非平稳噪声环境,能较好地检测耳语音的端点。由于耳语音信噪比较低,采用传统的算法进行耳语音端点检测存在正确率低、抗噪性能差等问题。提出了一种基于希尔伯特-黄变换瞬时能频值的耳语音端点检测的算法。运用希尔伯特-黄变换,分离出耳语音的瞬时幅值与频率,提取基于时间-能量-频率的特征参数瞬时能频值,利用该特征值对耳语音和噪声进行区分,进行端点检测。对700个信噪比为2～10dB的耳语音测试样本进行仿真实验,该算法检测的起点正确率与终点正确率均高于零能积法、熵法以及拟和特征法。实验表明,该算法适应于多种非平稳噪声环境,能较好地检测耳语音的端点。Because of the low SNR of the whispers, it is difficult to obtain both high accurate rates and good robustness with some traditional detection algorithms in whispered environment.An algorithm of whispered speech endpoint detection is presented which is based on Instantaneous Energy Frequency Value（IEFV） of Hilbert-Huang Transform（HHT）.This paper applies HHT to separate instantaneous amplitude and instantaneous frequency from whispers, and extracts IEFV,which is a temporal-amplitude-frequency character.Because IEFV can distinguish whispers form noise effectively, it is used as the feature for endpoint detection.The accurate rates of both initial and final of this algorithm are higher than the Zero-Energy-Product method, the Spectral Entropy method and the Fitting Characteristic method in the test with 700 samples at 2 - 10 dB SNR. As shown in the experiments, this algorithm can detect whispered speech endpoint accurately in various non-stable noisy backgrounds.

关 键 词：耳语音 希尔伯特-黄变换 瞬时能频值 端点检测 

分 类 号：TN912.34[电子电信—通信与信息系统]