【作者】 吉芳芳；

【导师】 张雪英；

【作者基本信息】 太原理工大学 ， 信号与信息处理， 2007， 硕士

【摘要】 语音识别是语音信号处理领域的研究热点，但由于其研究的复杂性，长期以来一直是一项难题，尤其是噪声环境下的非特定人语音识别。本文从一个典型的语音识别系统出发，介绍了语音识别的基本原理，讨论了几种常用的特征提取方法，尤其对过零率峰值幅度(ZCPA)特征提取作了较为详细的介绍。在此基础上提出用Laguerre滤波器对ZCPA特征提取前端处理进行改进的方法，并获得了具有优良抗噪性的识别结果。本文中用Laguerre网络实现的滤波器吸收了传统有限冲激响应(FIR)、无限冲激响应(IIR)滤波器的优点，既具有FIR滤波器的稳定性又具有IIR滤波器的长时记忆的特点和通阻带特性。其设计方法是在Laguerre滤波器与理想滤波器的频率响应的均方误差为最小的前提下，利用牛顿-拉夫逊法估算滤波器参数，然后由柯西-留数定理得出相对应的Laguerre系数以获得最优滤波器。通过实例设计了Laguerre滤波器，并与传统FIR和IIR滤波器的频率响应作了详细的比较，得出Laguerre滤波器有较小的滤波器长度，合适的线性相位和较少的通阻带波纹。缺点是计算复杂，但使用介绍的引理可降低其复杂性。接着将Laguerre滤波器用在ZCPA特征提取中代替原来的FIR滤波器，后端分别利用RBF网络和HMM训练和识别。实验结果表明利用Laguerre滤波器代替FIR滤波器进行特征提取，其识别率明显提高，而且抗噪性有很大改善。论文最后分析了Laguerre序列的频率弯折特性，并将小波变换的多分辨特性与之相结合得出基于Laguerre网络的频率弯折小波变换，对其实现结构作了详纽介绍，同时也说明难点所在。提出下一步工作是将频率弯折小波变换用于特征提取中，期望得到好的识别结果。更多还原

【Abstract】 Speech recognition has become a hotspot in the field of speech signal processing. But it is not easy to solve perfectly because of its complexity, especially for the speech recognition of Speaker Independent in noisy environment. This paper introduced the fundamental of speech recognition and discussed some commonly used feature extraction methods and specially analysed the ZCPA feature extraction based on a classical speech recognition system. Based on above conclusion, it presented a method which improved the front-end processing of ZCPA feature extraction, and got the better recognition rate which has excellent anti-noise properties.The filter realized by Laguerre network is a compromise between the FIR and IIR. It not only possessed the stability of FIR, but also had the good property of pass-band and stop-band of IIR, and achieved a long time memory. In this paper, Laguerre filter’s design approach was performed by evaluating the filter parameter employed Newton-Raphson method and corresponding Laguerre coefficients and obtaining optimum filter employed Canchy theorem when the minimum-mean-square-error of the frequency response between the Laguerre filter and the optimum filter was existed. It designed Laguerre filter and compared the frequency response of Laguerre filter with the frequency response of conventional FIR and IIR filters through experiment, educed Laguerre filter has small length and appropriate linear phase with less ripples in pass-band and stop-band. But the cost is the complex computation which can be reduced by the introduced lemma.Based on the better property of Laguerre filter, it used Laguerre filter replacing the traditional FIR filter in the ZCPA feature extraction, and employed RBF network and HMM to train and recognise in the back-end. The experiment results showed applying Laguerre filter replacing the traditional FIR filter in feature extraction would improved speech recognition rate and anti-noise properties.At last, it combined Laguerre transform and wavelet transform based on the frequency warped properties of Laguerre transform and multiresolution of wavelet transform, detailedly introduced its structure, simultaneously also explained what is the difficulty, and the further task is applying the frequency warped wavelet transform to the feature extraction, it’s expect to gain the better speech recognition result.更多还原