Page 67 - 《应用声学》2024年第6期
P. 67

第 43 卷 第 6 期                                                                       Vol. 43, No. 6
             2024 年 11 月                         Journal of Applied Acoustics                 November, 2024

             ⋄ 研究报告 ⋄



              结合改进DRSE-GCNN的电力调度语声识别模型                                                                    ∗





                             苌文涵      †   张云翔 顾 彬 相增辉 陈 轩 李霁轩


                                         (国网江苏省电力有限公司信息通信分公司           南京   210024)

                摘要:针对现有电力调度中语声识别方法存在的字识别错误率高和耗时长等问题,在分析语声识别技术的基
                础上,提出了一种改进的深度学习方法用于电力调度语声识别。将改进的深度残差收缩网络和改进的门控卷
                积神经网络相结合,通过改进的深度残差收缩网络提取有效特征,在通过堆叠改进的门控卷积神经网络来获
                取有效上下文信息。通过试验对所提方法的性能进行分析,验证其优越性。结果表明,所提方法与常规识别方
                法相比,在模型参数、字识别错误率和平均识别时间上均具有一定的优势,模型参数量为 6.48 M,字识别错误
                率为 2.87%,平均识别时间为 0.187 s。该研究为电力调度语言识别方法的发展提供了一定的参考。
                关键词:电力调度;语言识别;深度残差收缩网络;门控卷积神经网络;字识别错误率
                中图法分类号: TM933            文献标识码: A         文章编号: 1000-310X(2024)06-1243-07
                DOI: 10.11684/j.issn.1000-310X.2024.06.007



                       Combining improved DRSE-GCNN for power dispatching voice
                                                  recognition model


                             CHANG Wenhan       ZHANG Yunxiang      GU Bin   XIANG Zenghui

                                                 CHEN Xuan      LI Jixuan

                   (Information and Communication Branch of State Grid Jiangsu Electric Power Co., Ltd., Nanjing 210024, China)

                 Abstract: In order to solve the problems of high error rate and long time consuming in speech recognition in
                 power dispatching, an improved deep learning method is proposed for power dispatching language recognition
                 based on the analysis of speech recognition technology. Combining the improved deep residual shrinkage
                 network with the improved gated convolutional neural network, the effective features are extracted by the
                 improved deep residual shrinkage network, and the effective contextual information is obtained improved gated
                 convolutional neural network. The performance of the proposed method is analyzed through tests to verify its
                 superiority. The results indicate that, compared with conventional recognitionmethods, the proposed method
                 has certain advantages in model parameters, character recognition error rate, and average recognition time. The
                 model parameter quantity is 6.48 M, the character recognition error rate is 2.87%, and the average recognition
                 time is 0.187 s. This study provides a certain reference for the development of language recognition methods
                 for power dispatch.
                 Keywords: Power dispatch; Language recognition; Deep residual shrinkage network; Gated convolutional
                 neural network; Word recognition error rate


             2024-01-23 收稿; 2024-03-25 定稿
             国网科技项目 (DGF5687485)
             ∗
             作者简介: 苌文涵 (1991– ), 女, 安徽泗县人, 本科, 工程师, 研究方向: 电力调度通信, 图像语声信号处理。
             † 通信作者 E-mail: changwenhan2023@163.com
   62   63   64   65   66   67   68   69   70   71   72