【作者】 赵丽霞；

【导师】 赵欢；

【作者基本信息】 湖南大学 ， 计算机科学与技术， 2010， 硕士

【摘要】 语音端点检测的目的是从包含语音的一段信号中确定出语音的起点和终点,是语音信号处理的前端操作,在语音增强、语音编码、语音识别等领域得到广泛应用。语音端点检测方法有基于特征和基于模型两类,基于模型的方法比较复杂,对环境的适应能力差,而基于特征的方法相对简单且具有一定的抗噪能力,此方法要求找到某种能够区分语音和噪声的鲁棒性特征。本文针对基于特征的语音端点检测方法展开研究。针对基于谱熵的检测算法在低信噪比下鲁棒性差的缺点,提出一种新的基于距离熵的检测算法。该算法利用熵和倒谱系数的鲁棒性改变概率密度的计算方法,对经过预处理的带噪信号进行一系列运算得到每一点的倒谱系数,根据倒谱系数获得欧式距离,由欧式距离构造概率密度函数,由概率密度函数得到距离熵特征,最后利用距离熵采用双门限值进行语音和噪声的区分。本文还提出了一种基于支持向量机的多特征检测算法。基于支持向量机的检测算法对带噪信号分别求信噪比、修正过零率和AMMM三个特征,将三个特征组成一个特征矩阵,使用部分带噪信号对支持向量机进行训练,利用训练后的支持向量机自动区分语音和噪声。本文实验所使用的带噪信号由法国aurora2.0库的干净语音和Noisex92噪声库的噪声混合而成,并使用MATLAB工具进行仿真实验,实验结果表明,本文提出的两种端点检测算法具有一定的鲁棒性,在较低信噪比下仍能较好的区分语音和噪声。更多还原

【Abstract】 The purpose of voice activity detection (VAD) is detecting the beginning and ending points of speech from a signal which contains speech. As a pre-operation of speech signal processing, VAD is very important and has potential applications in the areas of speech enhancement, coding, identification and so on. VAD methods could be divided into two categories:feature-based and model-based. Model-based VAD method is complex and has poor adaptation to environment; Feature-based VAD method which requires finding some robust features to distinguish between voice and noise is relatively simple and has some anti-noise capability. This paper focuses on researching feature-based VAD method.Because of VAD method based on entropy can not work well in noisy environment, we propose a new algorithm based on distance entropy. This algorithm makes use of robustness of cepstral and entropy, change the calculate way of probability density function. We obtain cepstral coefficients of each speech point by a series of operations on noisy signal which have been pre-processed. We can get Euclidean distance according to cepstral coefficients, and then, we generate probability density function by Euclidean distance and construct distance entropy by way of probability density function. Finally, we can find useful parts of noisy speech by distance entropy.In addition, we propose another improved algorithm called VAD algorithm based on support vector machine by multi-feature. This algorithm extracts SNR, amending zero crossing rate and AMMM three characteristics from noisy signal, the three characteristics format a characteristic matrix. We employ parts of the noisy signal to train the support vector machine, set parameters, then support vector machine can distinguish noise from speech automatically.Noisy signals used in experiments are mixed by clean speech and noise. Clean speech come from French aurora2.0 Library, and noise come from Noisex92 noise library. Experiments tool is MATLAB. Simulation results prove that the two algorithms proposed in this paper perform well on anti-noise, they can work well in high-noisy environment.更多还原