基于小波分析的语音识别的研究

【作者】 张威；

【导师】 孟传良；

【作者基本信息】 贵州大学 ， 通信与信息系统， 2008， 硕士

【摘要】 语音识别技术的应用,本质上在于它能将输入的语音转化为语言代码,能够大幅度降低代码率,便于存储和传输,而且也容易被计算机或专用信息处理单元理解其含义,从而开发出更广泛的应用。例如,机器能听懂人类的自然语言。能够有效去除语音信号中的噪声是当今业界研究的热点问题,有很重要的理论价值和实用意义。本论文研究被噪声干扰的语音信号的去噪和识别课题。首先对语音信号和噪声的特性进行了分析;接着对语音识别系统的预处理、语音信号分析方法、特征提取、模板训练和模板匹配方法进行了论述;语音识别率的提高需要提取准确的语音特征参数,最好的办法就是对待识别语音进行降噪处理。本论文选取小波变换阈值去噪原理去除噪声。在对众多小波函数的分析中选择了sym8小波基和Heursure阈值选择规则,在‘sln’重调方法的前提下,分别采用硬阈值法、软阈值法和双变量阈值法,以及不同的小波分解层数进行了实验,得出采用双变量阈值法和5层尺度分解得到比较好的去噪效果和较小的信号损失的成果,对解决小波基选择和小波阈值选择的两个难点问题提供了一个可行的方法。更多还原

【Abstract】 The application of speech recognition technology allows the input speech signal to be changed into speech code. With the technology, not only the data of the speech, which is transferred and storied in code mode, is less than that in original way, but the speech code is easier processed by computer or other information process unit. Therefore, the speech recognition technology can be applied in many fields, for example, a machine can understand out language. Efficient speeches de-noise which is a research focus in IT is meaningful for real world and has high theoretical value.The theme is about de-noising of speech with noise and speech recognition. Firstly, the feature of speech signal and noise is introduced, and then the components of the speech recognition system, such as preprocessing, means of speech signal analysis, feature extraction, the training and the matching of speech template, are discussed. To increase the rate of speech recognition, the parameter of speech feature should be extracted accurately, signal de-noise is the best way to achieve the goal.The ’sym8’ wavelet and ’Heusure’ threshold rule are chosen. Under the ’sln’ readjustment method, hard, soft and double threshold are separately adopted in the experiments of different layer wavelet. The results of the experiment support 5 layers criterion decomposition with the double threshold, with which we can get good de-noise effect and reduce the lost of signal. And the study provide a effective method of wavelet and threshold selection更多还原