Abstract: In order to increase the speech signal’s efficiency and decrease the redundancy of the speech signal, the valid information of speech signal is often extracted by endpoint detection technology. However, the accuracy of traditional spectral entropy endpoint detection is decreased rapidly in low Signal-to-Noise condition. This paper proposes a method which is based on weighing threshold. Every spectral entropy of noise frame will be weighed with the initial noisy spectral entropy. The new weighed spectral entropy will be used as the refreshed threshold. At the same time, the spectral subtraction is used in the decision procedure to boost up the SNR and keep the robustness. The experiment results indicate that the accuracy of the speech endpoint detection has been improved in low SNR condition.

Key words: endpoint detection, spectral subtraction, spectral entropy

摘要： 为了进行有效的语音信号处理，降低语音信号的冗余度，通常采用端点检测技术来提取语音信号中的有效部分。而传统谱熵端点检测算法由于判定门限为固定值，其在低信噪比条件下检测性能急剧下降，提出了一种基于动态加权门限的检测方法，对每个判定的噪音帧的谱熵与无声段噪音谱熵进行加权平均，得到新的噪音谱熵作为更新后的门限值；在判定过程中引入谱减法提高信噪比，进一步降低噪声干扰。仿真实验结果证明，相对于传统谱熵端点检测方法，该方法在低信噪比的条件下仍然能够更为准确地检测到语音的端点。

关键词: 端点检测, 谱减法, 谱熵