基于字素-音素匹配的翻译机器人设计与研究  

Design and study of a translation robot based on phoneme-phoneme matching

作　　者：陈静 刘鹏 CHEN Jing;LIU Peng(Xi’an Siyuan University,Xi’an 710038,China;Xi’an Wanying Education Consulting Co.,Ltd.,Xi’an 710000,China)

机构地区：[1]西安思源学院,西安710038 [2]西安万鹰教育咨询有限公司,西安710000

出　　处：《自动化与仪器仪表》2025年第3期154-157,共4页Automation & Instrumentation

基　　金：横向课题《企业资料翻译技术服务》(NZ-H20240647)。

摘　　要：针对传统英语翻译机器人存在形音匹配准确率低,导致后续翻译效果不佳的问题,提出设计一个基于字素-音素匹配的翻译机器人可视化交互系统。首先,对字素-音素和形音匹配原理进行具体分析;然后采用基于隐马尔可夫模型(Hidden Markov Model,HMM)对目标语音文本信息进行抽取;最后通过双向长短时记忆神经网络(Bi-directional Long Short-Term Memory,BiLSTM)进行发音预测,最终获得一套完整的形音匹配规则。实验结果表明,提出的方法对英语词汇发音和字素发音的预测准确率分别为96.84%和97.25%,预测准确率明显高于传统的Seq2Seq预测方法和SVM预测方法,且将本方法应用到翻译机器人可视化交互系统中进行测试发现,本方法能够实现多个单词字素-音素的快速精准匹配,满足翻译机器人可视化交互需求,具备一定有效性。In view of the problem that the traditional English translation robot has low accuracy of shape and sound matching,which leads to poor subsequent translation effect,a visual interaction system based on phaceme-phoneme matching is proposed.Firstly,the principle of phoneme-phoneme and shape and tone matching is analyzed;then the target speech text information is extracted by hidden Markov model(Hidden Markov Model,HMM),and finally by bi-directional Bi-directional Long Short-Term Memory(BiLSTM) to obtain a complete set of shape and tone matching rules.The experimental results show that the proposed method of English vocabulary pronunciation and phoneme pronunciation prediction accuracy is 96.84% and 97.25% respectively,the prediction accuracy is significantly higher than the traditional Seq2Seq prediction method and SVM prediction method,and the method applied to the translation robot visualization interaction system found that the method can achieve multiple word phoneme-phoneme fast accurate matching,meet the requirements of translation robot visualization interaction,have certain effectiveness.

关 键 词：形音匹配 翻译机器人 HMM 双向LSTM 可视化交互 

分 类 号：TP392[自动化与计算机技术—计算机应用技术]