摘要
提出一种基于贝叶斯最小错误率的语音端点检测算法.说明短时平均幅度和短时平均过零率两种语音特征的选取方法,根据最小错误率的贝叶斯决策分别确定了合适的门限值,并将该方法在作者建立的语音数据库中进行了语音起点和终点测试.实验结果证明,与仅用幅度或能量特征的方法相比,本文方法能够提高语音端点检测的准确性.
A approach to speech endpoint detection is presented based on the Bayes minimum error probability. The feature of magnitude average and zero cross ratio average in short time are introduced. Then, an adaptive threshold is applied to detect the endpoint of speech by the Bayes minimum error probability policy decision. This method is tested on the speech database which is established by the author and the experimental results have proven its precision on speech endpoint detection compared with the method only using the feature of the magnitude or energy.
出处
《天津工业大学学报》
CAS
2008年第6期62-64,70,共4页
Journal of Tiangong University
关键词
语音端点检测
贝叶斯决策
埠时平均幅度
短时平均过零率
最小错误率
speech endpoint detection
Bayes policy decision
magnitude average in short time
zero cross ratio average in short time
minimum error probability