基于神经网络和卡尔曼滤波算法的说话人识别

Speaker recognition based on artificial neuron network and Kalman filter algorithm

首先从语音信号中提取出特征参数:线性预测倒谱系数(LPCC)和用小波包提取的小波特征参数(WPC);语音特征分类模型则选择多层前馈式神经网络(MBP网络),并将奇异值分解运用到扩展卡尔曼滤波(EKF)算法中作为神经网络的学习算法。仿真结果表明,小波特征参数具有良好的识别效果;同时采用改进后的扩展卡尔曼滤波(EKF)算法使人工神经网络具有更稳定、更准确的分类性能。

The LPCC coefficients and the wavelet packet coefficients are distilled from the speech. They are combined with the multi-BP network so that Kalman filter learning algorithms make the text-independent recognition. The experiment results show that the wavelet packet coefficients have good performance and the Kalman filter learning algorithms improves the capability of the neural network.