摘要
为提高噪声环境下语音端点检测的准确性,提出一种基于Mel倒谱距离顺序统计滤波的端点检测算法.该算法首先提取每帧语音信号的Mel频率倒谱系数,以前16帧估算背景噪声,计算每帧语音与背景噪声的倒谱距离;然后将当前帧前后相继若干帧的倒谱距离,经过一组顺序统计滤波器得到加权倒谱距离;最后根据各帧加权倒谱距离对输入语音进行分类.在TIMIT语音库上的实验结果表明,该方法在白噪声、粉噪声、汽车噪声和战斗机噪声等噪声环境下,均能得到理想的端点检测结果,且在低信噪比时依然有效.
To improve the accuracy of voice activity detection( VAD) under noisy environments,a VAD algorithm based on Mel frequency cepstrum coefficients( MFCC) distance with an order statistics filter( OSF) is proposed. First,the MFCC for each frame of the signals is extracted.Then,the background noise is estimated using the headmost sixteen frames. Finally,the MFCC cepstrum distance between each frame and the background noise is calculated. An order statistics filter is applied to a sequence of the estimated cepstrum distances to obtain the weighted cepstrum distance of each frame. The speech /non-speech classification is based on the weighted cepstrum distance. The experimental analysis carried out on the TIMIT speech corpus shows that the proposed algorithm is effective under white noise,pink noise,car noise,and fighter noise conditions even at low ratio of signal to noise.
出处
《中国科学院大学学报(中英文)》
CAS
CSCD
北大核心
2014年第4期524-529,共6页
Journal of University of Chinese Academy of Sciences
基金
国家自然科学基金(61005019
61273268
90920302)
北京市自然科学基金(KZ201110005005)资助