复杂真实环境下的调度电话转录算法研究  

Research on Dispatch Call Transcription Algorithm in Complex Realistic Environments

作　　者：詹丛茵 鲁工圆[1,2] 高辉 钱立 陈历泉 ZHAN Congyin;LU Gongyuan;GAO Hui;QIAN Li;CHEN Liquan(School of Transportation and Logistics,Southwest Jiaotong University,Chengdu 611756,Sichuan,China;National Joint Engineering Laboratory of Integrated and Intelligent Transportation,Southwest Jiaotong University,Chengdu 611756,Sichuan,China;Traffic Control Office,China Railway Guangzhou Group Co.,Ltd.,Guangzhou 510088,Guangdong,China)

机构地区：[1]西南交通大学交通运输与物流学院,四川成都611756 [2]西南交通大学综合交通运输国家地方联合工程实验室,四川成都611756 [3]中国铁路广州局集团有限公司调度所,广东广州510088

出　　处：《铁道运输与经济》2024年第4期83-93,100,共12页Railway Transport and Economy

基　　金：国家重点研发计划项目(2022YFB4300504);四川省自然科学基金项目(2022NSFSC0397);中国铁路广州局集团有限公司科技研究开发计划课题(KYL202301-0006)。

摘　　要：调度电话的应答是调度员日常工作的重要部分,也是事故回放审查的重要依据。为了提高事故分析及回放审查的效率,为调度指挥智能化打下基础,提出一种基于注意力机制及连接时序分类联合训练的调度电话转录算法,旨在提高复杂现实环境中调度电话的转录精确率及稳定性。通过在公开数据上添加调度大厅噪音来模拟现场环境,从而测试不同信噪比条件下算法的转录性能。在0 dB、1 dB、5 dB、10 dB的信噪比条件下测试转录错字率,结果分别是17.37%,16.48%,12.46%和9.69%,证明算法在强噪声环境下依旧能够保持稳定。算法包含半监督训练,实验表明,在小规模数据上,半监督训练相较于有监督训练可保证1%~2%的错字率降低。最后,算法在调度电话数据集上进行测试,并取得87.39%的关键字转录正确率。The handling of dispatch calls is an essential part of a dispatcher’s daily work,and serves as a crucial basis for accident review and investigation.To improve the accident analysis and playback review efficiency,and lay the foundation for the intelligent dispatching command establishment,this paper proposed a dispatch call transcription algorithm based on the joint training of attention mechanisms and connectionist temporal classification,aiming to improve the dispatch call transcription accuracy and transcription stability in complex realistic environments.By adding noise to public data to simulate complex on-site environments,the algorithm’s performance under different signal-to-noise ratio(SNR)conditions was tested.This paper tested the transcription error rates under SNR conditions of 0 dB,1 dB,5 dB,and 10 dB,and the results were 17.37%,16.48%,12.46%,and 9.69%,respectively,proving that the algorithm remains stable even in high-noise environments.Meanwhile,the algorithm includes semi-supervised training,and experiments showed that on a small-scale dataset,semi-supervised training can ensure a 1%-2%reduction in transcription error rate compared to supervised training.Finally,the algorithm was tested on the dispatch call dataset and achieved a keyword transcription accuracy of 87.39%,which can meet the professional requirements for dispatching commands.

关 键 词：铁路运输 调度电话转录 多任务训练 铁路调度指挥 半监督训练 自动语音识别 

分 类 号：U285.12[交通运输工程—交通信息工程及控制]