基于预测神经元模型的语音线性预测系数求解新方法

New Method to Solve the Speech Linear Prediction Coding Coefficients Based on Predictive Neuron Model

采用预测神经元作为语音信号线性预测模型的一种实现形式 ,可将线性预测系数的求解问题转化为预测神经元的训练问题 ,并运用 BP算法得到了神经元权值 (即线性预测系数 )的递推计算公式 .考虑到语音信号能量的不确定性 ,提出了运用相对预测误差能量作为收敛判断的参数 ,并按清音和浊音两种情况讨论了收敛判据 .由于利用预测神经元的迭代训练算法 ,理论上可以最大限度地挖掘语音样本中的相关性 ,因而可得到非常精确的线性预测系数 .计算结果表明 ,运用预测神经元方法所得到的线性预测系数 。

In this method, a linear prediction model of speech production is implemented by predictive neuron. The solving problem of the LPC coefficients is turned into a training problem of the predictive neuron. The iterative calculation formulas of the weights of neuron, i.e. LPC coefficients, are deduced based on the BP algorithm. On account of the uncertainty of speech energy, it is proposed that the relative prediction residual energy is used as the discriminative parameter of training convergence, and the convergence criterions of voiced phoneme and unvoiced phoneme should be discussed respectively. Because of utilizing the iterative training algorithm of predictive neuron, the correlation among speech samples can be extracted in the maximal degree theoretically. The calculation results indicate that the precision of method based on predictive neuron model is evidently higher than that of conventional Durbin algorithm and lattice algorithm.