基于小波变换的语音段起止端点检测算法

A Wavelet-Based Algorithm for Detecting the Beginning and End Points of Voice Segments

提出一种基于小波变换的语音段起止端点检测算法。传统的语音段起止端点检测算法大都是在时域上根据能量累积的大小判别语音段和噪声段,这些算法只是适用于高信噪比的情况。对于低信噪比的情况,这些算法往往要借助平均过零率等辅助特征。这样做不但增加了算法的复杂度,而且也未必奏效。单音节或浊辅音汉字的平均过零率与噪声的平均过零率大致相当。根据小波变换的特性,针对主要由白色噪声组成的噪声背景,提出一种新的语音段起止端点检测算法。这种算法根据白色噪声在小波变换域各个子带的平均能量变化平缓的特点判别语音段和噪声段。实验结果表明,算法即使在低信噪比的情况下也能正确判别语音段和噪声段。

A wavelet_based algorithm for detecting the beginning and end points of voice segments was proposed. The common_used algorithms are mostly based on their energy differences in time domain and therefore only suitable for the high SNR situation. For the low SNR, the common_used algorithms have to be strengthened by making use of some additional features such as average across_zero rates. This improvement will increase the algorithm complexity and still be possible to fail in some cases such as those Chinese characters with sonant consonants.A new wavelet_based algorithm has been proposed for the situation with white noise environment. White noise environment can remain unchanged across various subbands of wavelet transform domain and therefore can be detected and removed from voice signal. The experimental results show that this new algorithm can efficiently distinguish the beginning and end points of voice segments even in the low SNR environments.