预分浊音型的LPC基音提取法

本文介绍一种预分浊音型的LPC基音提取算法,对语言信号先用线性预测系数a₀和a₁的差值分出浊音区,然后只对浊音部分进行基音提取。提取基频时,数据率减半,用LPC的自相关方法产生8个预测系数的倒滤波器,倒滤波后的误差信号,用平均幅差函数（AMDF）方法提取基频,再线性插值,最后用非线性平滑滤波,并将所得结果和一个半自动的精确算法,以及简化倒滤波（SIFT）算法进行比较,说明我们提出的算法,对背景噪声40dB以下的连续语言是准确有效的。它避免了清音和无声间隙区的音调计算,且浊音和清音的判别比较准确。

A METHOD OF PITCH EXTRACTION BY LPC ANALYSIS WITH PREDETERMING VOICED SOUNDS

A method of pitch extraction by Linear Prediction Code (LPC) analysis with predeter-ming voiced sounds is presented in this paper. At first, the voiced parts in speech signal is identified by the difference a₀-a₁, where a₀=1 and a1 is the first coefficient of linear predictor, then only the voiced parts are processed to extract the pitch period. When pitch detecting, the data rate is compressed by half, the 8 coefficients of inverse filter are calculated by autocorrelation LPC method, and then for the residual signal, the output of the inverse filter, the pitch period is extracted by the Average Magnitude Difference Function (AMDF) method, and the curve of pitch period is linearly interpolated and nonlinearly filtered.
The results obtained by our method is compared with those obtained by a semi-automatic pitch detector and by SIFT method. It is shown that the algorithm presented in this paper is quite precise and effective for continuous speech with background noise below 40 dB. It is unnecessary toalculate the pitch periods in voiceless sounds and silence periods, and the classification of voiced or unvoiced sounds is rather precise.