语音信号的同态线性预测编码

Homomorphic Linear Predictive Coding of Speech Signals

本文首先证明普通LPC算法对于周期性脉冲串激励信号,在非基音同步的情况下是一种有偏估计算法。随后提出一种新的全极点模型估计算法——同态线性预测编码(HLPC),并证明,不论对于白噪声激励还是周期性脉冲串激励信号,由HLPC算法得到的估计都是无偏的。实验结果表明,对于用全极点模型人工合成的语音信号,HLPC的估计性能明显高于LPC;对于实际语音信号,HLPC算法在估计共振峰时也有明显的优点。本文还讨论了其它HLPC算法的优缺点和尚待解决的一些问题。值得指出,HLPC算法在任何以全极点模型(AR模型)为基础的估计问题中都有可能得到应用。

At first, we have proved in this paper that in the case of pitch-asynchronous analysis, the conventional LPC algorithms make biased estimations for the periodic pulse train excitation. Then we presenta new algorithm for the estimation of an all-pole model--Homomorphic Linear Predictive Coding(HLPC),which is an unbiased estimation algorithm, no matter whether the excitation is white noise or periodic pulse train The experiment results show that HLPC is obviously superior to LPC for artificial speech signals synthesized by an all-pole model. For real speech signal. HLPC is also distinguished for the estimation of formants. Moreover, we discuss the advantages and disadvantages and some problems to be solved for HLPC. HLPC may be applied to any estimation problems based on all-pole model (AR model).