论文部分内容阅读
近年来,深层神经网络(deep neural network,DNN)被成功应用于语音识别领域,成为一种很具发展潜力的语音识别模型。然而,由于其训练算法复杂度高,随着训练数据和网络规模增大,DNN模型训练将非常耗时。为提高DNN的训练效率,该文研究了基于多图形处理器(graph-ic processing unit,GPU)的DNN快速训练算法。在TIMIT数据集上的音素识别实验显示:在基本保证识别性能的前提下,优化后的DNN快速训练方法在4个GPU下训练速度相比单GPU有约3.3倍的提升。实验结果表明该快速训练方法可以显著提升DNN模型的训练速度。
In recent years, the deep neural network (DNN) has been successfully applied in the field of speech recognition and has become a promising speech recognition model. However, due to the complexity of its training algorithm, DNN model training will be very time-consuming as training data and network size increase. In order to improve the training efficiency of DNN, this paper studies DNN fast training algorithm based on graph-ic processing unit (GPU). The experiments of phoneme identification on TIMIT dataset show that under the premise of ensuring recognition performance, the optimized DNN training method can improve the training speed by about 3.3 times compared with single GPU under 4 GPUs. Experimental results show that the fast training method can significantly improve the training speed of DNN model.