一种基于优化小波神经网络的语音识别

Speech recognition using an optimized wavelet neural network

在以往的BP小波神经网络中,最常用的学习算法是BP算法,BP算法实质上就是梯度下降法,是一种局部搜索算法,梯度下降法使得网络极易陷入局部最小值,从而使得网络训练结果不尽人意,搜索成功概率低.取代传统的梯度下降法,利用粒子群算法对小波神经网络中的参数进行优化.然后利用基于粒子群优化(PSO)的小波神经网络进行抗噪声语音识别实验,仿真结果表明,与BP网络相比,PSO算法在迭代次数、函数逼近误差、网络性能方面均优于BP网络,系统的识别率也得到较大的提高.

The traditional BP wavelet neural networks usually adopt BP learning algorithm. BP algorithm is a gradient descent algorithm in essence, i.e. , a local search algorithm. The gradient descent algorithm is easy to fall into a local minimum, so the result of network training is unsatisfactory. Instead of the gradient descent algorithms, we use particle swarm optimization algorithm to train the parameters of the wavelet neural network. Then we use the PSO-WNN in speech recognition in noisy environment. Compared with the BP network, the simulation results show that the iterative number, function approximation errors and the network performance are highly improved than BP network. The recognition rate are also raised.