基于改进的CNN语音识别研究

针对卷积神经网络进行语音识别时识别率较低的问题,结合序列的最大子序列理论,把真实数据和预测数据看作两个序列并计算两者的最大子序列,再使用欧氏距离计算MSLoss损失函数.使用闵氏距离和神经网络反向更新时的参数,提出自适应卷积核ACKS算法,根据网络传播情况动态地改变卷积核大小,改善模型在不同阶段对数据特性的提取效果.设...

RESEARCH OF SPEECH RECOGNITION BASED ON IMPROVED CNN

Aiming at the problem of low recognition rate of speech recognition based on CNN,with the longest common subsequence theory,the real data and prediction data were regarded as two sequences,and the longest common subsequence of them was calculated.Then the MSLoss function was calculated by Euclidean distance.The ACKS algorithm was proposed combining the Minkowski distance and the parameters in the back updating of the neural network,which is able to dynamically change the size of the convolution kernel on the basis of network propagation conditions and improve the extraction of data features at different stages of the model.The improved network structure was designed,and the improved network was compared with the recurrent neural network and the long short-term memory neural network in the recognition rate and calculation time.The experimental results show that the improved model can reduce computing time by 2%and reduce error recognition rate by 3%.