小波分析在语音信号基音检测中的应用研究

【作者】 杨树功；

【导师】 张捷；

【作者基本信息】 西北工业大学 ， 通信与信息系统， 2005， 硕士

【摘要】 在语音信号数字处理的各个领域里,无论是语音分析与合成、语音压缩编码,还是语音识别和说话者确认等,准确可靠地检测语音信号的基音周期都是至关重要的任务,将直接影响到整个系统的性能。 本文旨在寻找一种鲁棒的基音周期检测算法。首先,在现代语音学取得的成果上,对语音的生成、声学特征,听觉功能进行分析,以把握语音信号波形的特点。其次,对较典型的几种语音基音检测方法,作了较系统的分析、探讨和比较。接着,较为详细地阐述了小波分析的基本理论,着重在小波函数、小波变换、多分辨分析和Mallat算法的研讨上。最后,将小波变换与归一化自相关、动态规划平滑等技术相结合,得到了一种新的基音周期检测算法。新算法的基本思想是:先对原始语音信号进行多级小波变换,将较高几个层次上的逼近信号进行加权求和处理,利用小波变换的带通性和去噪性,得到含丰富基音信息、周期性较强的合成信号(这一结果对含不同基音周期的任意语音段具有普适性):接着将该合成信号作为待处理信号,采用归一化自相关法检测基音周期,并对结果进行动态规划平滑处理。 用MATLAB在计算机上完成算法程序设计,进行仿真实验。实验结果表明,新算法估计基音准确性高,运算速度较快,稳定性好,对噪声具有较好的鲁棒性,充分吸收了自相关算法和小波变换算法的优点,有效地克服了自相关算法的分频和倍频现象,也比小波变换算法具有更强的抗噪能力,性能明显优于传统方法。更多还原

【Abstract】 In all fields of the digital processing of speech signals, for instance, speech analysis, speech synthesis, speech compression coding, and speech recognition with confirm by speaker, etc, detecting pitch period accurately and reliably is an essential task. It will influence the performance of the whole system seriously.This paper aims at looking for a kind of robust pitch detection algorithm. Based on the achievement of modern phonetics, the generation of speech, its acoustic characteristics and human’s sense of hearing are analyzed and characteristics of the waveform of speech signal are acquired. Then, a more systematic analysis and comparison among some typical methods of pitch detection is made. In addition, this paper describes the basic knowledge of wavelet theory in detail, studies the wavelet function, wavelet transform, Multi-resolution analysis, and Mallat algorithm. Finally, we combine wavelet transform, normal autocorrelation, and dynamic programming technology, propose a kind of new pitch detection algorithm. The main idea of the new algorithm is: Firstly, we carry on multistage wavelet transform to original speech signal, deal with the approaching signals on several higher levels by way of weighting and summing, get the synthetic signal with abundant pitch information and stronger periodicity (this result is general and right to the random section of speech including different pitch periods). Secondly, based on the synthetic signal, we adopt the normal autocorrelation method to detect pitch period and deal with the final data by dynamic programming technology.Finish designing program with MATLAB on the computer and carry on the simulation experiment. Experiment’s results show that new algorithm has high accuracy in pitch estimate, high speed in operation, good stability, and strong robustness to the noise. It combines the advantage of autocorrelation algorithm and wavelet algorithm, overcomes the phenomena of fractional frequency and double frequency in autocorrelation algorithm, and has better character of resisting the noise than wavelet algorithm. The performance of our new algorithm is obviously superior to the traditional algorithm.更多还原