一种噪声环境下的实时语音端点检测算法

语音识别中的端点检测要求对噪声有很强的鲁棒性。该文提出一种方法,综合采用了语音信号中的4个相互之间独立性强的特征-短时能量、倒谱距离、能量谱方差和能量-熵特征,有效地改进传统的基于单一语音特征方法的缺陷,在动态变化的噪声环境中,大大提高了端点检测对噪声的鲁棒性;为了克服分类回归树(CART)决策法的过度复杂性,引入一种新的5状态自动机进行快速决策,以保证算法的实时性能,并且能够提高端点检测的可靠性。通过各种实际噪声环境的测试,实验表明这一算法可以显著提高在低信噪比、噪声动态变化的各种环境下的端点检测性能。

A Robust Algorithm for Real-time Endpoint Detection in Noisy Environments

In speech recognition,the endpoint detection must be robust to noise.A method is presented in this paper,four speech features,e.g.short -time energy,cepstral distance,energy variance and energy -entropy,are taken into consideration.Because of the high independence of those four features,this method can adapt to various environments.The described algorithm not only uses four features but also introduces a5-states automation decision logic to increase the robustness in both low SNR and various noisy environments.The performance of the algorithm is evaluated by experiments in various noisy environments,and the performance of this endpoint detection algorithm is greatly improved.At the same time ,the proposed algorithm has a low complexity and is very suitable for real time mobile conditions.