面向电力线路巡检的语音指令识别系统研究和应用  

Research and Application of Voice Command Recognition System for Power Line Inspection

作　　者：张四维 武永泉[1] 秦涛 彭冲 赵彦杰[1] 焦良葆[2] Zhang Siwei;Wu Yongquan;Qin Tao;Peng Chong;Zhao Yanjie;Jiao Liangbao(State Grid Nanjing Power Supply Company,Nanjing 211899,China;Artificial Intelligence Industrial Technology Research Institute,Nanjing Institute of Technology,Nanjing 211167,China)

机构地区：[1]国网南京供电公司,南京211899 [2]南京工程学院人工智能产业技术研究院,南京211167

出　　处：《信息化研究》2021年第5期6-12,共7页INFORMATIZATION RESEARCH

基　　金：国家自然科学基金青年基金资助项目(No.61903183)。

摘　　要：为解决电力线路巡检时传统人工查询的低效和费时等问题,以及通用语音识别工具针对电力专业指令识别率低的问题,文章提出了面向电力线路巡检的语音指令识别系统。首先针对电力专业词汇,建立相应的基础语料库;在语音信号识别引擎建模中,基于电力专业指令的短时依赖性选择时延神经网络-隐马尔可夫模型(TDNN-HMM)构建特征提取网络和初级网络(STT);最后根据电力指令的专用语法结构和词汇库,提出了N元模型(N-gram)的指令纠错矫正方法,最终实现了低错误率的指令识别。实验结果表明,基于专用电力指令基础语料库的训练,TDNN-HMM识别网络,以及基于N-gram模型的指令纠错矫正方法均提升了识别准确度,所设计的专用识别引擎满足了工程实际需求。In order to solve the inefficiency and time-consuming problems of traditional manual query during power line inspection, and the general voice recognition tool for the problem of low recognition rate of power professional commands, a voice command recognition system for power line inspection is proposed. First, we establish the corresponding basic corpus for the electric power professional vocabulary;in the speech signal recognition engine modeling, based on the short-term dependence of the electric power professional instruction, we choose Time Delay Neural Networks-Hidden Markov Model(TDNN-HMM) to construct the feature extraction network and Speech To Text(STT) primary network;Finally, according to the special grammatical structure and vocabulary of power commands, an instruction error correction and correction method based on the N-gram model is proposed, which finally realizes instruction recognition with low error rate. The experimental results show that the training based on the basic corpus of special power instructions, TDNN-HMM recognition network, and instruction error correction method based on the N-gram model all improve the recognition accuracy, and the designed particular STT system can fulfill the demand of power line inspection.

关 键 词：智能电网 电力词汇 时延神经网络-隐马尔可夫模型 N元模型 声学模型 

分 类 号：TN912.3[电子电信—通信与信息系统]