Research on Speaker Verification Based on Cloud Computing Platform

At present, the accuracy of speaker verification is not high and the training speed is slow. In order to improve the accuracy of speaker verification, this paper uses transfer learning and score fusion technology. Experimental results show that with the knowledge learned from the VoxCelebl dataset, transfer learning reduced the equal error rate (ERR) by 0.67% on the aishell-l dataset. By fusing the convolutional neural network-based model CNN and the DenseNet-BiGRUbased model DR, the ERR is reduced above 0.63%. In the aspect of training speed improvement, this paper uses data parallel technology based on the cloud platform, and studies the impact of the communication window value of Downpour mechanism and starting multiple workers in each cloud node on training speed and accuracy. The experimental results show that starting multiple workers in each cloud node can effectively improve the training acceleration ratio. This method has a more significant effect on the model DR of lower communication overhead. At the same time, the Downpour mechanism can also improve the training acceleration ratio, and it has a better acceleration effect on the model CNN of higher communication overhead. In the case where the cloud platform has 3 nodes, the acceleration ratio of the DR model is 3.94, which exceeds the linear acceleration ratio. Meanwhile, the impact on the speaker verification accuracy is less than 1% compared with stand-alone training. The acceleration ratio of the CNN model is 2.69, which is close to the linear acceleration ratio.