论文部分内容阅读
本文对音素识别中时延神经网络提出若干改进训练方法并与原方法进行实验对比研究，发现通过采取如下措施可有效地增加时延神经网络的收敛速率：（1）误差反传法初训权值。（2）从单极性输出改为双极性输出。（3）改变能量函数使权值修正根据输出误差的大小而改变。（4）将反传误差修正权值从按时延帧取平均改为按层进行。这些措施使收敛时间从原来的23小时另25分减少到45分钟，收敛速率提高数十倍之多，而网络复杂度增加很少。
In this paper, several improved training methods are proposed for the time-delay neural network in phoneme recognition and compared with the original method. It is found that the convergence rate of the time-delay neural network can be effectively increased by adopting the following measures: (1) value. (2) From unipolar output to bipolar output. (3) change the energy function so that the weight correction according to the size of the output error change. (4) The anti-pass error correction weight from time-delay frame average to layer by layer. These measures reduced the convergence time from the original 23 hours to 25 minutes and the convergence rate increased by a factor of 10, with little increase in network complexity.