摘要
噪音谱的估计是谱相减方法中关键的一环。传统的噪声谱的估计是通过对输入语音作语音检测,区分出纯噪声段,根据噪声段的频谱估计出噪声谱。该方法的准确性局限于语音检测算法的性能,在信噪比较低时,性能下降很快。近年来人们提出了多种不用语音检测的噪声估计方法,这些方法不区分语音和非语音段,在每一帧都进行噪声谱的更新。评估了几种无语音检测的噪音估计方法,比较了它们用于谱相减时在语音识别中的性能,提出了一种新的基于能量聚类的无语音检测噪音估计方法,通过实验验证了它的优良性能。
In conventional noise estimation, voice activity detection (VAD) is usually used, and noise statistics are updated during speech gaps. VAD is essential in these methods. When the SNR is low, the performance deteriorates quickly. Recently, some noise estimation techniques without VAD are proposed. These techniques continuously update noise estimates whether speech is present or not. Several such approaches are evaluated and their performance in speech recognition is compared. A new energy clustenng based on noise estimation technique without VAD is also proposed. Experiments show its eminent performance.
出处
《计算机工程与设计》
CSCD
2003年第10期40-42,共3页
Computer Engineering and Design
基金
上海市科学技术委员会基础研究项目基金资助项目(01JC14033)
美国贝尔实验室上海分部的资助项目
