一种噪声环境下的语音端点检测方法

New method for detecting endpoint of speech in noise environment

利用短时过零率来检测清音,用短时能量来检测浊音,两者相配合便实现了信号信噪比较大情况下的端点检测。但是在信噪比较小的环境下,这两种方法便失去了作用。为了能在噪声环境下准确地检测出语音信号的端点,根据对含噪语音在时频域中的研究,提出了一种基于Matching pursuits时频分解算法的语音端点检测方法。该方法使用Matching pursuits算法对含噪信号进行分解,然后再对信号进行魏格纳变换,可以完全去除信号的魏格纳交叉干扰项,使得语音信号和噪声信号在时频平面上具有较直观明显的魏格纳能量分布,利用这个特点再进行端点检测,实验结果表明,该方法能在信噪比较低的情况下,准确地检测出语音信号的端点。

By studying noisy speech in the time-frequency domain, a new speech detection method based on decomposition with matching pursuits is proposeck With matching pursuits, the decomposed time-frequency energy of the signal has a clear distribution in the time-frequency domain. Using the speech characteristics, endpoints can be differentiated by setting a threshold value. Compared with the conventional method that uses zero crossing ratio to detect the surd and uses short-time energy to detect the sonant, although the real-time capability is sacrificed, the accuracy of endpoint is greatly improved in noisy environments.