语音识别论文(长文档排版).docx

摘 要 语音作为人机交互的新方法，因其灵活性、便捷性备受关注。语音识别在工业、金融、商业、文化、教育通信、医疗、家庭服务、消费电子产品等诸多领域都有广阔的应用前景，已经成为计算机领域非常活跃的课题之一。本论文从语音识别的过程出发做了以下几方面研究： 1) 简明介绍了语音识别技术的发展历程和趋势，分析了当前语音识别面临的主要问题。 2) 研究语音信号处理的关键技术，包括预处理、端点检测、特征参数提取等。端点检测作为语音识别的前端预处理部分，其正确性直接影响语音识别的准确率。本文讨论了基于双门限算法语音信号端点检测算法。通过实验仿真可看出该算法的有效性。在特征提取方面，着重研究了LPCC和MFCC两种特征参数。由于特征向量都是高维的，需要对数据作降维处理，这里介绍了SVD、PCA、ICA和特征选择四种方法。 3) 为了改善小样本情况下语音识别的效果，引入基于支持向量机的语音识别方法，讨论了支持向量机多类分类策略和训练算法。 4) 分析现实需求，设计一个基于微软的Speech SDK 5.1的语音家庭辅助护理系统，完成系统并做出了改进总结。 关键词：语音识别，端点检测，特征提取，支持向量机，语音辅助护理系统  Abstract As a new method for human-computer interaction, the voice system has drawn extensive attentions because of its flexibility and convenience. Boasting brilliant prospect of application in spheres including industry, finance, business, culture, educational communication, medicine, home care, and e-products, etc, the voice system has become one of the hot subjects for the computer circle. This paper mainly carries out researches on the following aspects in the speech recognition process: 1）This paper briefly introduces the developing process and tendency of speech recognition technology, and analyzes the major problems facing the speech recognition development. 2)This paper researches on the key technologies of voice signal processing, including preprocessing, end point detection, and feature parameter extraction, etc. As a part of preprocessing of the front end of speech recognition, the correctness of end point detection directly impacts on the accuracy of speech recognition. This paper discusses the detection method of voice signal end point based on double-threshold algorithm. Through simulative experiments, we can see the effectiveness of this algorithm. In regard to the feature extraction, this paper gives the emphasis to two feature parameters which are LPCC and MFCC. As feature vectors are all high-dimensional, the dimensionality reduction of data should be processed. Altogether 4 methods including SVD, PCA, ICA and feature selection are introduced