基于隐马尔可夫模型的能量参数预测量化算法

HMM-Based Prediction and Quantization Algorithms for Energy Parameters

为了充分利用能量与线性预测编码(Linear prediction coding,LPC)系数之间的相关性,提高能量参数量化效率,提出了一种基于隐马尔可夫模型(Hidden Markov model,HMM)的能量参数预测量化算法。通过适当假设,使用HMM模拟能量参数和LPC系数之间的相关性,其中离散化后的能量参数组成隐状态序列,量化后的LPC系数组成可观测序列。然后利用HMM预测每一超帧中的能量参数的变化轨迹,并根据预测出的能量轨迹对预测残差进行分模式矢量量化(Mode-based vector quantization,MBQ)。仿真实验中能量参数量化后的平均失真为2.668 dB,与线性预测量化算法相比下降了14.0%,表明本文算法通过利用能量参数与LPC系数的相关性,能够有效地提高能量参数量化效率。

To use the correlation between energy parameters and linear prediction coding（LPC） coefficients,and to quantize the energy parameters more efficiently,hidden Markov model（HMM） based prediction and quantization algorithms are proposed.HMM is used to model the correlation between the energy and the LPC coefficients under appropriate assumptions.In HMM,the discretized energy parameters constitute hidden state sequences and the quantized LPC coefficients constitute observation sequences.HMM is used to predict the energy contour of each super frame,and then mode-based vector quantization（MBQ） is applied to quantize the energy prediction errors according to the predicted energy contour.Experimental result shows that the average quantization distortion is 2.668 dB,which is reduced by 14.0% comparing with linear prediction and quantization algorithms.It implies that the proposed algorithms can improve the energy quantization efficiency by using the correlation between energy parameters and LPC coefficients.