非平稳噪声环境下基于谐波能量的语音检测

A Speech Endpoint Detection Method Based on Consonance Energy

语音端点检测的鲁棒性,对于构建实际语音识别系统具有重要的意义。谐波成分是语音信号的一个基本特点,为此提出了一种基于谐波成分能量的端点检测算法。通过sobel算子计算窄带语谱图的方向场,通过Gabor滤波增强谐波区域,通过门限方法得到二值化图,去除方向大于45度和依赖度低的点,得到连续的水平方向的带状分布,即谐波分布区域,求取谐波分布区域内的能量,以此作为门限判决的特征。实验结果表明,在不同信噪比、多种非平稳噪声环境下都能够达到较好的语音检出效果。其优点为,不需要噪声的先验知识,充分利用了语音在频率域和时间域的相关性,适应于各种非平稳复杂噪声。

The robustness of VAD（voice activity detection） is crucial to the construction of a practical automatic speech recognition system. This paper presents a new VAD algorithm based on the energy of consonance, which serves as a basic feature of speech. The consonance region is attained through the following stages： orientation estimation, consonance enhancement, binarisation and post - process via Sobel operator, Gabor filter, threshold classifier and pruning respectively. The pruning is performed mainly through discarding those regions with orientation greater than 45 degree or low dependency. The consonance energy, sum of the consonance regions energy, is then used as the feature of the threshold decision argument of VAD. The experimental results show that the proposed algorithm holds the following advantages： independent of the prior information, fully utilizing the correlation of frequency and time domain of speech, applicable for various complex non - stable noises.