面向铁路领域的多方言免切换语音识别方法  

A Multi-Dialect Switching-Free Speech Recognition Method for Railway

作　　者：杨立鹏 胡从刚 陈华龙 韩可可 刘峰 张志科 YANG Lipeng;HU Conggang;CHEN Huaong;HAN Keke;LIU Feng;ZHANG Zhike(Institute of Computing Technology,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China;Passenger Transport Department,China State Railway Group Co.,Ltd.,Beijing 100844,China)

机构地区：[1]中国铁道科学研究院集团有限公司电子计算技术研究所,北京100081 [2]中国国家铁路集团有限公司客运部,北京100844

出　　处：《中国铁路》2025年第1期30-39,共10页China Railway

基　　金：中国铁道科学研究院集团有限公司科研开发基金项目(2023YJ132)。

摘　　要：随着铁路智能客服系统的持续发展,现有语音识别模型的准确率已达到较高水平,但面对铁路领域术语和多样化方言场景,其语音识别效果仍然较差,因此提出1种融合铁路领域知识的多方言免切换语音识别方法。基于RepVGG网络模型构建方言语种识别器,以获取语种信息;对Transformer语音识别模型进行改进,通过在编码器中融合注意力机制的语种残差模块,并在解码器中嵌入语种信息,以实现多方言免切换功能;基于LSTM网络模型在铁路文本语料库上训练铁路领域专用的语言模型,并将其与改进的Transformer模型进行融合,以提升对铁路术语的识别准确率;在自建数据集上对所提方法进行实验验证,结果表明:所提方法在粤语、四川话和普通话上的识别准确率均超过90%,且有效提升对铁路领域术语的识别性能,具有一定应用价值。With the continuous development of intelligent railway passenger service systems,the accuracy of current speech recognition models has reached a high level.Nevertheless,their performance remains undesirable when dealing with railway-specific terminology and diverse dialect scenarios.Therefore,a new multi-dialect switch-free speech recognition method that incorporates railway-related knowledge is proposed.A dialect recognizer was developed using the RepVGG network model to acquire dialect information.The Transformer speech recognition model was enhanced by integrating a dialect residual module with an attention mechanism in the encoder and embedding the dialect information into the decoder,enabling a multi-dialect switching-free capability.Using the LSTM network model,a railway-specific language model was trained on a railway text corpus and integrated with the improved Transformer model,boosting the accuracy of railway terminology recognition.This method was tested on a custom dataset,and the result shows that it achieves over 90%recognition accuracy in Cantonese,Sichuanese,and Mandarin,significantly improves railway terminology recognition,and has great potential for application.

关 键 词：铁路领域语言模型 多方言 语音识别 RepVGG TRANSFORMER LSTM 

分 类 号：U29-39[交通运输工程—交通运输规划与管理] TP18[交通运输工程—道路与铁道工程] TP391.4[自动化与计算机技术—控制理论与控制工程]