卷积神经网络声学模型的结构优化和加速计算

将卷积神经网络(convolutional neural networks,CNN)声学模型应用于中文大词表连续电话语音识别任务中,分析了卷积层数、滤波器参数等变量对CNN模型性能的影响,最终在中文电话语音识别测试中,CNN模型相比传统的全连接神经网络模型取得了识别字错误率1.2％的下降.由于卷积结构的复杂性,常规的神经网络加速方法如定点量化和SSE指令加速等方法对卷积运算的加速效率较低.针对这种情况,对卷积结构进行了优化,提出了2种卷积矢量化方法:权值矩阵矢量化和输入矩阵矢量化对卷积运算进行改善.结果表明,输入矩阵矢量化方法的加速效率更高,结合激活函数后移的策略,使得卷积运算速度提升了8.9倍.

Structure optimization and computing acceleration for convolutional neural network acoustic models

In this paper, we apply convolutional neural network acoustic models to a continuous telephone speech recogni-tion task with large vocabulary in chinses. We studied the influences of the number of convolution layers and filter parame-ters on the model performance. Finally, in the Chinese telephone speech recognition test, the CNN model obtained an abso-lute 1.2％ decrease in terms of word error rate compared with the traditional fully connected neural network model. Due to the complexity of the convolution structure, conventional neural network acceleration methods such as fixed-point quantiza-tion and SSE instruction have lower acceleration efficiency for convolution operations. In this case, we have optimized the convolution structure and proposed two kinds of convolution vectorization method:weight matrix vectorization and input ma-trix vectorization to accelerate the convolution operation. Results show that the input matrix vetorization method is more effi-cient and an 8.9 times acceleration with no performance degradation is obtained together with the activations retroposition strategy.