本文叙述一种语言信号处理的新方法——快速递归梯型算法。用算符分解的办法导出这种算法自回归(AR)模型的基本公式及其归一化方案。说明导出参数快速收敛和跟踪的性能;研究如何应用它的基本参数(似然变量、反射系数和预测误差的协方差等)来检测音调脉冲,爆破音和其他快速过渡特性。同时给出了合理的掩蔽参数,设计一个有效的自适应门限的局部极值选定算法,从而消除从原参数产生的多余信息和误差。这一音调检测算法已应用到一个可变数据率(平均800bits/s)的声码器的模拟设计中,得到相当满意的结果。 另外这种算法要求极少的存贮和简单的算术运算,便于硬件实现,尤宜VLSI的实现。 This article describes a new method of speech signal processing - fast recursive ladder algorithm. The method of operator decomposition is used to derive the basic formula and normalization scheme of AR algorithm. Describe the performance of deriving parameters for fast convergence and tracking; examine how to apply their basic parameters (likelihood variables, covariances of reflection coefficients and prediction errors, etc.) to detect pitch pulses, plosives, and other fast transition characteristics. At the same time, a reasonable concealment parameter is given and an effective local threshold selection algorithm of adaptive threshold is designed to eliminate redundant information and errors from the original parameters. This pitch detection algorithm has been applied to a variable data rate (average 800bits / s) vocoder simulation design, get quite satisfactory results. In addition this algorithm requires very little storage and simple arithmetic operations, easy to implement hardware, especially VLSI implementation.