【作者】 乔永凤；

【导师】 马建芬；

【作者基本信息】 太原理工大学 ， 计算机应用技术， 2007， 硕士

【摘要】 盲源分离是指在不知道源信号分布和混合系统的情况下，仅根据观测到的混合信号恢复源信号的过程。由于盲源分离无需知道信号的先验信息，从而在信号处理领域得到广泛的应用，语音盲分离更是因为其实用性成为其中研究的热点。语音分离技术对计算机听觉、语音识别等方面的研究具有重大意义，同时高质量的语音通信、助听器、电话远程会议系统也都得益于此，因此，语音盲分离的研究具有非常重要的理论价值和应用价值。本文就语音信号盲分离进行了研究，主要有以下几点：(1)提出了一种基于最小互信息和MLPs、RBF神经网络的语音信号盲分离算法。该算法是在infomax算法的基础上，结合前向型神经网络的信息后向传播，优化目标函数，在分离过程中尽可能的提取信号的独立分量，把信号分离出来。同时比较了两种网络在该算法下对分离语音信号的优化性能，仿真实验表明，该算法能够成功的分离混合语音信号，并且RBF网络比MLPs网络性能更好。(2)提出对于语音信号盲分离而言最佳的时—频分析窗结构。在基于时—频分析的盲分离中，窗函数的长度和形状对分离语音信号的性能有着重要的影响。本文实验仿真并比较了在不同窗函数和窗长下分离语音信号的性能，实验表明对于语音这种短时平稳信号，窗长选择为256时可取得最佳的分离效果。(3)针对语音的短时平稳性，详细分析了语音帧长和互信息之间的关系，语音间统计独立性随语音帧长的减小而增强。在帧长小于100ms时，语音信号间的互信息量显著增强。由此指出，尽管盲源分离算法能够以批处理方式分离语音信号，但用于实时语音环境中分离信号还存在一定的局限性。更多还原

【Abstract】 Blind source separation is a process of recovery source signals only according to observed mixed signals. It finds broad application owing not to require the priori knowledge of signals and the blind separation of speech signals becomes research hotspot because of its practicability. The research of blind speech signals separation is important for computer hearing and speech recognition, and high-quality speech communication, aid-hearing and telephony remote conference system can be profit from it. Therefore, the research of blind speech signals separation has very important value of theoretic and application.This paper is researched the Blind speech signals separation. The uppermost contributions are as follows:1. This paper presents an algorithm of blind speech signals separation based on the minimization of the mutual information and MLPs, RBF neural network. The algorithm is using the infomax algorithm, and combining the information backpropagation theory of neural network, optimizing objective function, making the yield components that are as independent as possible during the course of the separating the mixing signals. This paper compared the optimum performance of separated speech signals between two neural network, the experiment indicate that the algorithm can be successful separate the mixing speech signals, and the RBF is better than MLPs. 2. The best window structure of time-frequence for blind speech signals separation has been put forward in this paper. Due to the Length and shape of multiple Windows faction have important influence for the performance of separated speech signals; this paper experimented and compared it. The conclusion reveals that where the window length is 256, the separated signals can be attain the best effect for the quasi-stationary nature and inherent correlation of speech over the temporal short term.3. On the temporal short term of speech signals, we detailed analyses the relationship for frame size and mutual information. The statistically independent increase for speech as the frame size decreases, where the frame sizes less than 100 ms, the mutual information value dramatic increase exhibited by the speech signals. The results indicate that although algorithm of blind source separation is suitable for application with speech in batch techniques possessing substantial data, it is inevitably less reliable for realistic audio environments that require a real-time approach to separation.更多还原