贵阳方言工厂指令识别  

Identification of Factory Instruction in Guiyang Dialect

作　　者：李凯飞 徐凌桦[1] LI Kaifei;XU Linghua(The Electrical Engineering College,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学电气工程学院,贵阳550025

出　　处：《微处理机》2022年第1期48-51,共4页Microprocessors

基　　金：国家自然科学基金(61861007);2020年贵州大学混合式课程建设项目“计算机控制技术”(2020030)。

摘　　要：针对贵阳工厂环境下口头任务对接缺乏依据性、出现事故难于追责的问题,引入深度学习模型改善贵阳方言工厂指令识别效果。自制贵阳方言工厂指令数据集,搭建指令识别系统,依次训练六种模型,其中包括拥有9层隐藏层的深度神经网络。在同一测试集下,系统随训练的进行逐渐提升性能,在DNN模型下识别错误率降至最低,远低于单音素模型识别错误率。对比不同测试集识别错误率,分析噪声对识别性能的干扰。实验表明DNN模型下带噪测试集错误率比纯净测试集高出不到3%,证明DNN模型具有更为优良的鲁棒性。In view of the lack of basis for oral task handover in Guiyang factory environment, and the difficulty in finding responsibility for accidents, a deep learning model is introduced to improve the factory instruction recognition effect of Guiyang dialect. The instruction data set of Guiyang dialect factory is made, and the instruction recognition system is built. 6 models are trained in turn, including the deep neural network with 9 hidden layers. Under the same test set, the performance of the system is gradually improved with the training, and the recognition error rate in DNN model is reduced to the lowest, far lower than that in monophone model. Comparing the recognition error rate of different test sets, the interference of noise on recognition performance is analyzed. Experiments show that the error rate of noisy test set under DNN model is less than 3% higher than that of pure test set, which proves that DNN model has better robustness.

关 键 词：语音识别 工厂指令 深度神经网络 贵阳方言 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TN912.34[自动化与计算机技术—控制科学与工程]