一种新型前向神经网络用于汉语四声识别

Chinese Tones Recognition Based on a New Neural Network

提出了一种新型的前向神经网络，并研究了其在语音识别中的应用．该神经网络为只含一层非线性隐层的前向神经网络，以线性的输出层代替一般ＢＰ网络的非线性输出层，可以更准确地、范围更大地完成非线性函数估值功能．该神经网络采用了包括反向传播算法（ＢＰ）及最小均方算法（ＬＭＳ）的混合算法进行训练，可以减少落入局部最小点的概率以及提高收敛速度．利用该神经网络可以很好地完成汉语四声识别功能．

A new kind of feed forward neural network and its application in speech recognition are described in this paper.The neural network has one nonlinear hidden layer,with linear output layer to replace the output of BP neural network,which has better ability of approximating a function.A hybrid algorithm combined with the back propagation and least mean square method can be used for training,which will reduce the probability of falling local minimal and improve the convergent speed.The experiment shows that it is very effective for the Chinese tones recognition.