时域和频域特征相融合的语音端点检测新方法  

A novel speech activity detection algorithm based on the fusion of time domain and frequency domain features

作　　者：刘欢 王骏[1] 林其光 王士同[1] 

机构地区：[1]江南大学数字媒体学院,无锡214122 [2]无锡百互科技有限公司,无锡214125

出　　处：《江苏科技大学学报（自然科学版）》2017年第1期73-78,共6页Journal of Jiangsu University of Science and Technology:Natural Science Edition

基　　金：国家自然科学基金资助项目(61300151);江苏省自然科学基金资助项目(BK20130155);江苏省高校自然科学研究项目(13KJB520001);科技部科技型中小企业技术创新基金(14C26213201061)

摘　　要：为了提高语音端点检测的适应性和鲁棒性,提出一种时域和频域特征相融合的语音端点检测新方法.在对语音信号进行预处理的基础上,对每一帧分别提取调和性、清晰度和周期性这3个时域或频域特征,使用主成分分析进行特征融合,并采用双门限法得到语音端点的候选集合.在此基础上通过支持向量机对候选集合中的端点进行判断得到最终结果.仿真实验表明:相对于传统的语音端点检测算法、时域和频域特征相融合的语音端点检测新算法提高了语音端点检测的正确率,有效降低了误测率和漏检率,具有更好的适应性和鲁棒性,对不同噪声背景的信号都有较好的检测能力.In order to improve the adaptability and robustness of speech activity detection,a novel algorithm for speech activity detection（SAD） is proposed based on the integration of time domain and frequency domain features. In the proposed method,three features,i. e. harmonicity,clarity,periodicity are extracted and combined together with principal component analysis. The candidates of the endpoints are detected by double-threshold method. SVM is utilized to determine the final set of endpoints based on the candidates. Experimental results indicate that the proposed SAD method is effective and provides superior and consistent performance across various noise and distortion levels.

关 键 词：特征融合 特征提取 支持向量机 语音端点检测 主成分分析 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]