基于语谱图的语音端点检测算法  

A Speech Endpoint Detection Algorithm Based on Spectrogram

作　　者：陈向民[1] 张军[1] 韦岗[1] 

机构地区：[1]华南理工大学电子与信息工程学院,广东广州510640

出　　处：《电声技术》2006年第4期46-49,共4页Audio Engineering

摘　　要：利用语音在语谱图中表现出的不同特征,提出了一种基于语谱图的语音端点检测算法。首先利用基音频率检测的原理在语谱图矩阵中搜索浊音段,然后计算出浊音段的信噪比,再根据信噪比和语谱图矩阵中浊音段的峰值进行完整的端点检测。因多数突发噪声并没有稳定的频率或者频率不在人的基音频率范围内,因此,该算法能够很好地抑制突发噪声的干扰,实验结果表明,在信噪比为10dB以上时该算法能够准确检测出语音的端点位置。In this paper, a new algorithm based on spectrogram for speech endpoint detection is presented, The new algorithm employs pitch detection method to find out the voiced speech segments from the spectrogram matrix, then the SNR in each voiced speech segment is calculated. Finally, the detection with SNR and peak value of each segment is completed. Experimental results show that the new algorithm can work well with the outburst nois- es and find out the endpoint of each speech segment accurately and robustly when SNR is above 10 dB.

关 键 词：端点检测 语谱图 清浊音 

分 类 号：TN912[电子电信—通信与信息系统]