摘要
针对低信噪比、非平稳噪声环境下的语音端点检测,提出了一种基于语音/噪声的信源系统动力学特性差异,通过分析信号递归度变化,设定双门限判定语音端点的方法。和传统的能量法、倒谱距离测度法比较,准确率较高。为语音特征提取和识别研究提供了新的途径。
In order to perform speech endpoint detection under extreme noisy and nonstationary conditions, a new algorithm based on the difference of underlying dynamics between speech and noise was proposed. Recthrence rate was taken as the quantification parameter to measure the variation of such difference and double decision thresholds were applied in the deci- sion framework. Comparing with some traditional methods such as short-time energy and cepstral distances, this algorithm shows lower wrong decision rate. This method provides a new way for feature extraction and speech recognition.
出处
《通信学报》
EI
CSCD
北大核心
2007年第1期35-39,共5页
Journal on Communications
基金
国家重点基础研究发展计划("973"计划)基金资助项目(2005CB724303)~~
关键词
语音端点检测
递归图
递归度分析
特征提取
speech endpoint detection
recurrence plot
recurrence rate analysis
feature extraction
