基于多特征的语音端点检测技术研究  

Study on Endpoint Detection Technologies of Speech Signals based on Multiple Characteristics

作　　者：何彬[1] 柳平[1] 王琦[1] 程行甫[1] 韩林呈[1] 

机构地区：[1]装甲兵工程学院信息工程系,北京100072

出　　处：《通信技术》2010年第11期139-141,共3页Communications Technology

摘　　要：针对传统的端点检测技术,如基于能量、过零率等方法,在低信噪比噪声环境下检测性能急剧下降的问题,根据汉语语音发音的特点,提出了一种新的检测方法,该方法结合了Mel频率倒谱系数(MFCC)和能量、过零率、频带方差等多个语音特征。基于多特征融合的模糊判决二次搜索端点检测方法,能有效减少清音、拖尾音的截断,提高端点检测的精度,并对噪声环境具有一定的自适应性。实验结果表明,即使在低信噪比条件下,该方法仍具有较高的准确性。Traditional voice activity detection（VAD）algorithms are based on speech properties,such as temporal energy variations and zero-pass ratio,and their efficiency would decrease sharply with low signal-to-noise-ratio（SNR）.Based on the particular characters of Mandarin Language a novel robust algorithm in combination of MFCC and spectral characteristics is proposed in this paper.This algorithm relies on fuzzy estimation and executes dual search,thus has excellent noise immunity.The experimental results show that even with a lower SNR,the detection accuracy is still high.

关 键 词：端点检测 模糊判决 噪声自适应 多特征融合 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]