混合声音信号辨别的并行化方法的研究与实现

发布时间：2018-04-08 14:47

本文选题：声源辨别　切入点：声音信号分离　出处：《江苏科技大学》2017年硕士论文

【摘要】：人们听到的声音往往都是由多个声音混合而成的,如何从混合的声音信号中快速而准确的分辨出感兴趣的声音信号,一直是研究的热点。传统的方法可以进行简单的声源辨别,但是当涉及到大数据量的声音信号处理时,影响了其应用的实时性和准确性。随着人工智能时代的到来,以深度学习和GPU并行计算为代表的新技术为大数据量的声音信号处理提供了解决思路,为此本文设计了混合声音信号辨别的并行化方法,并开展了以下工作:1.分析了国内外混合声音信号的研究现状以及发展趋势,以混合声音信号为切入点,学习了混合声音信号辨别和GPU并行计算的相关知识,并研究了混合声音信号分离以及声源辨别的常用方法。2.对混合声音信号进行去均值和白化等预处理,选取基于负熵的Fast-ICA算法进行混合声音信号分离,通过分析混合声音信号分离过程寻找制约其快速分离的原因,并利用GPU并行化进行加速改进。3.对分离后的声音信号进行多特征值提取,并将提取出的特征值进行融合组成复合特征值,再进行声源辨别。在辨别过程中,由于传统神经网络存在学习能力不足的问题,针对这个缺陷,引入了基于深度信念网络(DBN)的声源辨别模型,以提升混合声音信号辨别的准确率。4.由于要进行大数据量的声音信号处理,并且声音信号在处理过程中同时又具有方法一致、独立性强的特点,于是采用GPU并行化方法分别对基于负熵的Fast ICA算法、特征值提取和深度信念网络模型的训练过程等操作进行优化,提高了混合声音信号辨别方法的处理效率。通过仿真和实验验证,利用GPU并行化对混合声音信号的辨别方法进行优化改进,提高了混合声音信号分离和辨别的效率,满足了实时性要求。同时,采用基于多特征值融合的复合特征值作为输入数据和基于深度信念网络的声源辨别模型,提升了混合声音信号辨别的准确率。
[Abstract]:The sound that people hear is often composed of multiple sounds. How to quickly and accurately distinguish the interesting sound signal from the mixed sound signal has always been a hot research topic.The traditional method can distinguish the sound source easily, but it affects the real time and the accuracy of the application when dealing with the sound signal processing of the large amount of data.With the arrival of the era of artificial intelligence, the new technology, represented by deep learning and GPU parallel computing, provides a solution for the sound signal processing of large amount of data. In this paper, a parallelization method for the discrimination of mixed sound signals is designed.And carried out the following work: 1.This paper analyzes the research status and development trend of mixed sound signal at home and abroad. Taking mixed sound signal as the starting point, we study the related knowledge of mixed sound signal discrimination and GPU parallel computing.The common methods of mixed sound signal separation and sound source discrimination. 2.The mixed sound signal is pretreated with de-mean and whitening, and the Fast-ICA algorithm based on negative entropy is selected to separate the mixed sound signal. The reason for the fast separation of mixed sound signal is found by analyzing the separation process of mixed sound signal.And using GPU parallelization to accelerate the improvement. 3.The separated sound signal is extracted with multiple eigenvalues, and the extracted eigenvalues are fused to form composite eigenvalues, and then sound source identification is carried out.Due to the deficiency of learning ability in traditional neural networks, a sound source discrimination model based on deep belief network (DBN) is introduced to improve the accuracy of mixed sound signal discrimination.The extraction of eigenvalues and the training process of the depth belief network model are optimized to improve the processing efficiency of the mixed sound signal identification method.Through simulation and experimental verification, the GPU parallelization is used to optimize and improve the discrimination method of mixed sound signals, which improves the efficiency of separation and discrimination of mixed sound signals and meets the real-time requirements.At the same time, the composite eigenvalue based on multi-eigenvalue fusion is used as input data and the sound source discrimination model based on deep belief network is used to improve the accuracy of mixed sound signal identification.
【学位授予单位】：江苏科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3

【参考文献】