基于时频增强和谱熵的语音端点检测技术  

Speech Endpoint Detection Based on Speech Time-Frequency Enhancement and Spectral Entropy

作　　者：武传艳[1] 范影乐[1] 

机构地区：[1]杭州电子科技大学自动化研究所,浙江杭州310018

出　　处：《杭州电子科技大学学报（自然科学版）》2005年第1期18-21,共4页Journal of Hangzhou Dianzi University：Natural Sciences

基　　金：国家自然科学基金资助(60302027);浙江省自然科学基金资助项目(602127);浙江省教育厅科研资助项目(20030620)

摘　　要：该文提出了基于时频增强和谱熵的语音端点检测算法。该算法对带噪语音在频域利用谱减法去除宽带加性噪声,在时域去除由谱减带来的残差噪声从而对语音进行了增强。对增强后的语音利用谱熵特征进行端点检测。实验结果表明,该算法快速有效,具有较强的抗噪能力,特别适合低信噪比的语音端点检测。In this paper, an improved method of speech endpoint detection based on time-frequency enhancement and spectral entropy is proposed. Under lower SNR, conventional methods can't perform well in endpoint detection. In this paper, the noise speech is enhanced using spectral subtraction method in spectral domain to remove the additive noises. In time domain, a weighting function built by short-time energy and zero-cross-rate is used to remove the music noise produced by the spectral subtraction. Entropy based method is used to detect the endpoints. Experiment results show that it outperforms the entropy based without enhancement endpoint detection under lower SNR.

关 键 词：谱熵 谱减法 语音端点检测 时频增强 

分 类 号：TN912.35[电子电信—通信与信息系统]