基于小波分析的语音端点检测算法研究.doc

基于小波分析的语音端点检测算法研究

摘 要
语音端点检测是语音识别中至关重要的技术。无论军用还是民用，语音端点检测都有着广泛的应用。在低信噪比的环境中进行精确的端点检测比较困难，尤其是在无声段或者发音前后。本文讨论了几种常用的端点检测方法，并提出两种基于小波分析的端点检测，并在此基础上描述了基于这两种算法的语音端点检测综合算法，从而实现对语音信号精确端点检测的方法。
文中首先介绍了几种常见的语音端点检测方法如短时能量与过零率，隐马尔可夫等。这些方法在静音环境下，当噪声较小或噪声相对单一时可以取的较好的检测结果，但在语音环境较恶劣、信噪比较低时，检测的结果下降较快，难以让人满意。为此本文引入了小波变换作为分析工具。接下来论文讨论了小波变换的原理及在语音识别系统中的应用。
论文分别提出了两种基于小波系数的语音端点检测方法，并对其实验结果进行了比较。第一种方法是子带平均能量方差用于语音端点检测，该方法利用噪声的分类及特点，以及它与语音信号的差别，在小波分析的基础上，对每一子带的平均能量进行方差分析，从而区分出语音段。该方法具有快速、简单和准确率高的特点。第二种方法是小波系数方差用于语音端点检测，语音信号是统计自相似的随机过程，它的统计特性在时域内不随波形的扩充或压缩而变化。根据这一特性为识别语音与背景噪声建立一个理想的贝叶斯两层分类器，以每一子带内的小波系数作为比较参数，从而进行分类计算。最后根据概率的大小得到端点检测的结果。该方法具有适用范围广、准确率高的特点，而算法相对比前一方法要复杂。
论文在讨论了前两种方法的优缺点、分析实验结果后，提出一种揉合两种方法，以发挥各自优点的新方法。实验表明该方法发挥了以上两种方法的特点具有很好的检测结果。

关键词：端点检测，小波变换，系数方差，子带能量
STUDY OF SPEECH ENDPOINT DETECTION ALGORITHM BASED ON THE WAVELET ANALYSIS
ABSTRACT
Speech endpoint detection is a key technology for speech recognition. It is widely used in not only military usage but also civilian usage. It is difficult to exactly detect endpoint under low SNR, especially in silence segment or before pronouncing or after pronouncing. This paper discussed several kinds of commonly used endpoint detection methods, and proposed two endpoint detection means based on the wavelet transform, and based on this integrated speech end-point detection algorithm, thereby the method of exact speech signal endpoint detection can be obtained.
This paper first introduced some kinds of commonly used speech endpoint detection methods, such as short-time energy and zero-crossing rate, HMM etc. Using these methods, the result is better under silence environment, less noisy or relatively single noisy environment, but under bad, low SNR environment, the decline of the result is fast, and the result is not satisfying. So this article presented using wavelet transform as analysis tool. Next, we discussed the principle of wavelet transform and its application in speech recognition system.
This paper proposed two kinds of speech endpoint detection methods based on wavelet coefficients respectively, and compared the two experimental results. The first method is sub-band average-energy variance used in speech endpoint detection. Based on wavelet analysis, utilizing the classification , character of the noise, and the difference between speech signal and noise, this method made variance analysis on the average energy of each sub-band to distinguish speech endpoint. This method has the feature of fast, simpleness and high exact rate. The second method is using the variance of the wavelet coefficients to detect speech endpoint. Speech signal is a random process with statistical similarity to itself, its statistical feature doesn’t vary with the expansion or compression of the waveform in time field. According to this feature, we establish an ideal Bayes classification models with two levels for recognizing speech and background noise. The classification calculation is based on considering wavelet coefficients variance of each sub-band as comparing parameter. At last, the result of endpoint detection is obtained through comparing their variance. This method has the feature of wide range for applicability and high accuracy, but the algorithm is more complex than the former method.
After discussing the merits and shortcomings of the former two methods and analyzing the experimental results, this paper presented a new method which combined the merits of the former two methods. The experiment indicated that this new method has good detection result for integrating the features of the above two methods.

KEY WORDS：speech endpoint detection，wavelet transform，parameter variance，subband-energy
目 录
第一章 绪 论 1
1.1概述 1
1.1.1 语音识别简介 1
1.1.2 端点检测在语音识别系统中的地位和作用 3
1.1.3 国内外研究现状 5
1.2几种常用的端点检测方法 7
1.2.1 短时能量及过零率 8
1.2.2 熵函数 10
1.2.3 LPC倒谱特征 11
1.2.4 隐马尔可夫(HMM) 13
1.3 课题研究背景 14
1.4论文内容安排 16
第二章 小波分析理论 18
2.1概述及特点 18
2.2小波分析与傅立叶分析的比较 19
2.3小波分析的基本理论 24
2.4 小波分析在语音处理中的应用 30
2.5 小结 33
第三章 子带平均能量方差用于语..