基于谱减与自适应子带对数能熵积的端点检测

Endpoint detection based on spectral subtraction and logarithmic energy entropy product of adaptive sub-bands

针对低信噪比环境下语音端点检测准确率低、鲁棒性差,提出了一种将谱减降噪和自适应子带对数能熵积相结合的语音端点检测算法。首先利用改进的多窗谱估计谱减法提升语音信号质量,再以自适应子带对数能熵积这一新的语音特征参数为阈值,使用动态阈值双门限检测法进行语音端点检测。实验结果表明,该算法针对低信噪比语音信号具有更好的准确性和鲁棒性。

Aiming at the problems of low accuracy and poor robustness of speech endpoint detection in low SNR environment,a speech endpoint detection algorithm combining spectral noise reduction and the product of logarithmic energy entropy of adaptive sub-bands was proposed.The improved multi-window spectral subtraction method was used to improve the quality of speech signals.A new speech feature parameter,the logarithmic energy entropy product of adaptive sub-band,was used as the threshold,and the dynamic threshold double-threshold detection method was used to detect speech endpoints.Experimental results show that the proposed algorithm has better accuracy and robustness for low SNR speech signals.