【作者】 蔡魁杰；

【导师】 沈继红；

【作者基本信息】 哈尔滨工程大学 ， 应用数学， 2007， 硕士

【摘要】 语音识别在工业、军事、商业、银行、服务、医疗、日常生活等很多领域有着广泛的应用前景。本文将支持向量机理论方法应用于语音识别中两个关键技术——端点检测和声韵母分离，取得了良好的效果。支持向量机(support vector machine，SVM)是数据挖掘中的一项新技术，是借助于最优化方法解决机器学习问题的新工具。它最初于20世纪90年代由Vapnik提出，近年来在其理论研究和算法实现方面都取得了突破性的进展，开始成为克服“维数灾难”和“过学习”等传统困难的有力手段。对于支持向量机的端点检测，本文提出了基于C-支持向量机的端点检测技术，解决了传统端点检测方法中需要人为设定阈值的繁琐和不准确性，并且可以用此方法提取语音识别研究中任意感兴趣的特征语音段。对于支持向量机的声韵母分离，传统声韵母分离方法一般需要人为预先设定阈值，这需要大量的试验分析和数据统计。本文提出了基于C-支持向量机的声韵母分离技术，不需预先人为设定阈值。本文研究了选用不同输入特征和不同惩罚参数情况下的支持向量机的分类能力，并且针对训练样本极其贫乏情况下的输出判断向量采取“侵蚀”的后期处理方法，提高了分类准确度。更多还原

【Abstract】 The speech recognition has a great application prospect in many domains such as the industry, the military, the trade service, the bank service, the medical service, the daily life and so on. Using the support vector machine theory in solving two pivotal techniques in speech recognition, we get a fine effect in endpoint detection and initial/final segmentation.Support vector machine (SVM) is a new method in data mining, and is a new tool for solving machine learning problems in the virtue of optimization methods. It was first proposed by Vapnik in 1990s, and made great progresses in theory research and algorithms application recent years, and is becoming a powerful method to overcome traditional difficulties such as "dimension disaster" and "excessive learning".For endpoint detection based on support vector machine, a new method based on C-support vector machine is proposed in this paper to solve the problem of ado and inaccurateness brought by doors initialization in traditional endpoint detection methods, and this method can also be used to detect and pick out speech segment in which people have special interests.For initial/final segmentation based on support vector machine, traditional initial/final segmentation methods need doors to be initialized first and this needs a great many experimentations and data analysis. We propose a method of initial/final segmentation based on support vector machine which needs no doors to be initialized before segmentation.Researching on the correlation between the classification ability and input character and punishing parameter, and in the condition of extremely lacking of training data, we propose an "eroding method" to process the output of the support vector machine, and so improve the classification precision.更多还原