低信噪比语音信号的经验模态分解与端点检测

摘要

In order to improve the accuracy of speech recognition and synthesis in the low SNR environment,the empirical mode decomposition and the spectral subtraction improved by multiple window spectrum estimation are used for speech signal reconstruction.And then the Teager energy operator and zero crossing rate are used to detect the speech signal after noise reduction.A speech endpoint detection algorithm based on the empirical mode decomposition and the improved spectral subtraction is proposed.Results show that,under the background of white noise,pink noise and Babble noise,the endpoint detection accuracy of this algorithm is higher than that of the algorithm based on the empirical mode decomposition and the improved double threshold speech endpoint detection method and that its detection accuracy decreases with the decreasing SNR,and decreases less than that of the algorithm based on the empirical mode decomposition and improved double threshold speech endpoint detection method,realizing the endpoint detection of speech signals in a low SNR environment.%为了提高在低信噪比环境下的语音识别和合成正确率,文中采用经验模态分解和多窗谱估计改进谱减法对语音信号进行了降噪重构,利用Teager能量算子和过零率对降噪后的语音信号进行了端点检测,给出了一种基于经验模态分解和改进谱减法相结合的语音端点检测算法.结果表明:该算法在白噪声、粉红噪声和Babble噪声背景下的端点检测正确率优于基于经验模态分解和改进双门限语音端点检测法;该算法端点检测正确率随信噪比的下降而降低,且降低幅度小于基于经验模态分解和改进双门限语音端点检测法,实现了低信噪比环境下语音信号端点检测.