基于多任务训练的用户登入语音识别模型仿真  

Simulation of User Login Speech Recognition Model Based on Multi-Task Training

作　　者：江官星[1] 付悦[1] JIANG Guan-xing;FU Yue(Science and Technology College of NCHU,Nanchang Jiangxi 332020,China)

机构地区：[1]南昌航空大学科技学院,江西南昌332020

出　　处：《计算机仿真》2022年第9期190-194,共5页Computer Simulation

摘　　要：传统用户登入语音识别模型的泛化性能较差,导致语音识别精度不理想。为解决上述问题,构建基于多任务训练的用户登入语音识别模型。利用循环神经网络(Recurrent Neural Network, RNN)的数据处理能力,将多任务学习(Multi-task learning, MTL)应用在循环神经网络中,采用共享隐层学习的方式并行训练多个任务,获取更多共享特征,完成多任务学习,提高循环神经网络泛化性能,构建基于MTL-RNN的语音识别模型,将用户登入连贯语音信息作为模型输入,结合多任务学习结构,通过用户身份、情感和性别的分类输出,实现用户登入语音识别。实验结果表明,上述模型具备较高语音识别准确率,语音识别非加权平均召回率较高,说明引入多任务学习可增强上述模型的语音识别的泛化能力,优化识别精度。Traditionally, the generalization ability of the traditional user login speech recognition model is low, so the accuracy of voice recognition is always not ideal. In the paper, a model of voice recognition for user login based on multi-task training was constructed. According to the data processing ability of Recurrent Neural Network(RNN),we applied multi-task learning(MTL) to RNN. We trained multiple tasks in parallel by using the method of sharing hidden-layer learning, thus getting more shared features. After finishing multi-task learning, we improved the generalization ability of RNN. Moreover, we built the model of voice recognition based on MTL-RNN and took the coherent voice information of user login as model input. On the basis of the multi-task learning structure, we used the classification output of user ID,emotion and gender to achieve the voice recognition of user login. Experimental results show that this model has high accuracy of voice recognition and high unweighted average recall. Therefore, it indicates that multi-task learning can enhance the generalization ability of voice recognition and optimize the recognition accuracy of the model.

关 键 词：多任务训练 用户登入 语音识别 循环神经网络 多任务学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]