改进支持向量机在说话人识别中的应用-信号与信息处理专业论文.docxVIP

摘 摘 要 I万方数据 I 万方数据 摘 要 将支持向量机(SVM)模型应用在说话认识别技术领域，是当前的一个研究热点。 SVM 是一种区分性模型，适合处理小样本的分类问题，具有很强的分类能力。但在 实际环境中，文本无关的说话人识别系统需要较大的语音数据，将 SVM 直接用于训 练模型时，会产生训练速度慢、算法复杂而难以实现和分类结果易受核函数参数影响 等问题。本文采用高斯聚类(GMM)方法对大量语音数据进行分类，利用粒子群算法 (PSO)对 SVM 的核函数参数进行优化，有效改善了说话人识别系统的识别性能。针 对标准 PSO 算法容易陷入局部最优的缺点，提出一种改进的 PSO 算法。本文主要工 作如下： 1、前端特征参数提取部分：详细分析了语音端点检测技术、特征选择和特征变 换技术。本文提出了谱熵端点检测算法，相对于传统的端点检测算法，即使在信噪比 较低的情况下，也能准确的检测出语音信号的始末点。此外，对不同组合的特征参数 进行分析，并对选择的合适的特征参数进行高斯聚类，减少 SVM 的训练量。 2、讨论了两种说话人识别方法：基于 PSO 优化 SVM 的说话人识别和改进的 PSO 优化 SVM 的说话人识别。利用 PSO 对 SVM 的惩罚系数和核函数参数进行优化，相 对于标准的 SVM，说话人识别系统的识别率得到了提高。针对 PSO 算法的缺点，对 其进行改进，使改进后的算法加快了收敛速度，提高系统识别能力。最后通过 Matlab 仿真，实验结果表明改进的 PSO 的识别系统有较高的识别率。 关键词：说话人识别；文本无关；谱熵；支持向量机；粒子群算法 II II 万方数据 Abstract In recent years, the model of Support Vector Machine (SVM) which applies on the field of speaker recognition is a hot spot of research field. SVM is a discriminative model with better capability of classification, which is suitable for processing small samples decision. But in practical applications, the system of text-independent speaker recognition need lots of speech, so SVM has some problems of training slowing, difficult to realize with complex arithmetic and the kernel function easily effect on the results of classification, which it deal with a great deal of training data directly. This thesis adopts a method of Gaussian Mixture Model (GMM) for clustering to deal with a large number of speech data, and uses Particle Swarm Optimization (PSO) to optimize the kernel function parameters of SVM for better performance of the Speaker recognition system. Aiming at the shortcoming of PSO which was easily relapsed into local extremum, an improved PSO was proposed. The main works are listed as the following: Front end feature extraction part.: It analyzes the end point detection, feature selection and feature transformation. This paper proposes a spectral entropy endpoint detection algorithm which compared with the traditional endpoint detection algorithm, it can d