基于噪声估计的改进能量熵语音端点检测算法

An Improved Energy-entropy Algorithm for Speech Endpoint Detection Based on Noise Estimation

针对传统能量熵的短时能量与子带谱熵容易受噪声环境影响,低信噪比下端点检测性能下降的问题,提出一种基于噪声估计的改进能量熵语音端点检测算法。首先对语音进行噪声估计并以此计算语音存在概率;然后利用估计的噪声能量修正短时能量,用语音存在概率作为加权系数优化子带谱熵,并将两者结合生成改进的能量熵;最后给出基于噪声估计的动态门限以及实时的端点检测策略。实验结果表明,在信噪比5 dB、0 dB的多种噪声环境中,基于噪声估计的改进能量熵端点检测算法相比传统能量熵算法与改进子带能谱比算法,检测正确率平均提升7%。

To solve the problem that the short-time energy and sub-band spectral entropy of traditional energy-entropy are sensitive to noise environment,and the performance of endpoint detection is degraded under low signal-to-noise ratio(SNR),an improved energy-entropy algorithm applied to endpoint detection based on noise estimation is proposed.Firstly,the noise is estimated and the speech presence probability is calculated.Then,the estimated noise energy is used to correct the short-time energy,and the speech presence probability is used as the weighting coefficient to optimize the sub-band spectral entropy,and the two are combined to generate the improved energy-entropy.Finally,a dynamic threshold based on noise estimation and a real-time endpoint detection strategy are given.Experimental results indicate that the algorithm based on noise estimation improves the detection accuracy by 7%on average compared with the traditional energy-entropy algorithm and the improved sub-band energy spectrum ratio algorithm under various noise environments with SNR of 5 dB and 0 dB.