【作者】 张康杰；

【导师】 赵欢；

【作者基本信息】 湖南大学 ， 计算机系统结构， 2007， 硕士

【摘要】 语音信号的基音周期是描述语音信号激励源的重要特征参数之一。它在语音编码、语音合成、语音信号分离、语音识别和说话人识别等方面有着广泛的应用。但由于基音周期本身固有的特性,目前还没有一种能适应任何人,任何应用和任何环境的基音检测算法。经典的语音信号基音检测的方法有:自相关函数法、平均幅度差函数法、倒谱法、时变傅里叶变换法、小波变换法、简单逆滤波追踪法等,自相关函数法需要中心削波处理,阀值难以确定,因此受噪声的影响较大,抗噪性能较差;后面三种方法效果较好,但都需要大量的乘法运算,计算量大,它们都不适合使用在运算和存储条件受到限制嵌入式实时处理设备中。根据语音信号的基音周期范围有限和周期相对稳定的特点,本文改进了可变长平均幅度差函数法(LV-AMDF),提出一种自适应幅度差法检测基音周期。它在语音非稳定段通过简单的谷值点评选机制,筛选当前谷值点以及历史谷值点,得到较精确的基音周期;在语音稳定段依据历史谷值点缩短语音段的比较范围,减少计算代价。本文还改进了浊音起止点检测算法,使浊音起止点的定位更精确。实验证明,该方法在不同的信噪比环境下有效地降低了半周期和倍周期点的发生率。更多还原

【Abstract】 Pitch period is one of the important parameter in the analysis and synthesis of speech signals. Pitch period information is used in various applications such as speech coding, pitch synchronous speech analysis and synthesis and speech recognition. However the task of the estimating pitch is very difficult because of the pitch’s properties. Therefore, no one algorithm that has been developed so far performs perfectly for different speakers, different applications and different environmental condition.There are many traditional pitch detections, such as: AUTOC, AMDF, CEP, STFT, Wavelet Transform, SIFT and so on. AUTOC needs the center clipping, and its noise immunity is low because of the difficulty for finding out the threshold. The last three perform better, but they are complicated for they need too much multiplication,and they are not suitable for the real-time imbedded equipments,that have shortage for their limited CPU and memory resources.According to the limitation and the stability of the speech pitch, the paper improves the LV-AMDF method, and initiates an adaptive speech pitch determination method. In the unstable speech, it can acquire the exact speech pitch according to the trough in the current window and the windows before by the simple trough-select system, in the stable speech this method can reduce the times of the speech pitch comparison, decreacing the computing cost. This paper also improves the method, which can acquire more accurate point for searching the beginning and ending point of the sonant. Simulations showed that the pitch extracting results of the improved method reduce the rate of the half and double speech pitch point in the speech in different SNR environment.更多还原