【作者】 修国浩；

【导师】 杨鼎才；

【作者基本信息】 燕山大学 ， 电路与系统， 2004， 硕士

【摘要】 语音识别技术是信息技术领域的重要发展方向之一，目前其面临的一个重要挑战就是如何提高噪声环境下的语音识别率。特征提取作为语音识别的第一步，其性能对整个系统性能——语音识别率——具有至关重要的影响。因此，本文以提高系统在噪声环境下的语音识别率为目标，以提取抗噪声的语音特征参数为研究重点，研究了噪声环境下具有鲁棒性的语音识别系统。本文在深入理解语音识别基本原理的基础上，首先，介绍了几种被广泛应用的语音特征参数提取方法。其次，详细探讨了非平稳随机信号的时频分析方法——维格纳分布，从语音信号的时变特性出发，充分利用维格纳分布的优秀特性，把其应用于语音特征提取中，并与语音信号的同态处理方法相结合，提取出两组新的特征参数，即基于维格纳分布的语音倒谱参数WD-MFCC和基于对称相关函数的语音倒谱参数WV-MFCC。同时还得到基于维格纳分布的语谱图。最后，深入研究了隐马尔可夫模型在语音识别中的应用，把本文提出的两组语音特征参数和先前介绍的几种特征参数分别应用于以该模型为识别分类器的语音识别系统中，仿真并分析了噪声环境下利用各种语音特征时该语音识别系统的识别性能。仿真实验结果表明，采用本文提出的两组新的特征参数可以有效地提高系统性能。更多还原

【Abstract】 Speech recognition is one of main branches in the information and technology field. How to improve the robustness of a recognizer in presence of background noise has been a vital difficulty. In speech recognition, the first step is the extraction of speech features, whose capability are crucial to the performance of the whole speech recognition system. Therefore, this paper aims at improving the performance of speech recognition system in noisy environment, focuses on robust speech feature coefficients extraction, studies robust speech recognition system in noise.Based on deeply comprehension in the fundamentals of speech recognition, some kinds of widely-used speech feature coefficients are introduced firstly. Secondly, the time-frequency analysis method of nonstationary random signal is discussed, which is named Wigner Distribution (WD). Based on the time-varying character of speech, this paper makes full use of the excellent characters in WD and applies it to speech processing. Then WD is combined with homomorphic processing technique to compute two kinds of feature coefficients, which are cepstral coefficients based on WD, named WD-MFCC, and cepstral coefficients based on symmetrical correlation function, named WV-MFCC. At the same time, a spectrogram is derived from WD of speech. Lastly, the application of Hidden Markov Model (HMM) to speech recognition is deeply studied. The two kinds of coefficients are applied to a speech recognition system which employs HMM as the recognizer, as well as the previously introduced feature coefficients. What’s more, the last part of this paper has simulated and analyzed the robustness of the speech recognition system in noise when applying different speech features. The simulation results shows that the new feature coefficients proposed in this paper can significantly improve the robustness of speech recognition system.更多还原