摘要
语音识别系统的实用化,需要对噪声有很强的鲁棒性,而噪声环境下的端点检测对整个识别系统性能起着关键的作用。提出一种基于语音学知识的两级起点检测方法,其中第一级选取短时能零比和短时谱幅作为初检特征,并采取自适应门限,第二级根据语音起点能量变化和语音性持续时间进行起点的确定。实验结果表明该方法在常见噪声环境下鲁棒性较好,且适于实时应用。
While speech recognition system is put into use,it must be robust to noise.The endpoint detection in noisy background plays an important role in the whole recognition system.A kind of method is presented in this paper which is based on phonetic knowledge and includes two-stage starting-point detection.Short-time EZQ(Energy Zero Quotient)and short-time spectra amplitude are adopted as initiative detection features in the first stage;the accurate starting-point is determined according to the variation of endpoint phonetic energy and the time lasting of the phonetic features in the second stage.The experiments show that its better robustness can be obtained in common noisy environments and its real-time application.
出处
《电声技术》
北大核心
2004年第5期51-54,共4页
Audio Engineering
关键词
语音识别
端点检测
起点检测
短时能零比
短时谱幅
speech recognition
phonetic
endpoint detection
starting-point detection
short-time EZQ
short-time spectra amplitude