基于端到端的英语翻译器语音自动纠正系统研究  

Research on end to end speech automatic correction system for English translators

作　　者：乔万俊 赵庆 QIAO Wanjun;ZHAO Qing(Shaanxi Institute of Technology,Xi’an,710300,China;Great Wall Holding Group,Baoding 071000,China)

机构地区：[1]陕西国防工业职业技术学院,西安710300 [2]长城控股集团,河北保定071000

出　　处：《自动化与仪器仪表》2023年第3期240-244,共5页Automation & Instrumentation

基　　金：汉考国际2021年科研基金项目《“一带一路”背景下<工业机器人技术基础及应用>中文+教学资源建设研究》(CTI2021ZB05);2022年度陕西省汉语国际教育专项课题《“中文+职业教育”背景下高职工科类教师专业发展研究》(2022HZ1457)。

摘　　要：针对英语翻译器语音自动翻译纠正准确率的问题,结合卷积神经网络(Convolutional Neural Networks, CNN)和Transformer模型,提出基于CNN+Transformer端到端的自动翻译纠正模型。模型通过利用CNN模型平移和旋转不变性的特点,以缓解语音信号的多变性,并结合Transformer模型的多头自注意力机制兼顾语音特征和距离特征帧之间的依赖关系,进而提高模型的并行计算能力,实现了高效、准确的英语翻译器语音自动翻译纠正。最后,通过在CCMT2019-BSTC数据集上进行仿真,验证了所提CNN+Transformer模型的有效性。仿真结果表明,所提CNN+Transformer模型具有良好的语音自动翻译效果,BLEU值为19.23,参数规模和训练时长分别为18.99 M和5.4 min,相较于当下常用端到端翻译模型seq2seq模型、级联翻译模型NMT模型和cascade模型,具有更好的翻译效果,且具有一定的有效性和优越性,可用于实际英语翻译器语音自动翻译。Aiming at the problem of poor automatic speech translation effect of English translators, this paper proposes an end-to-end automatic translation model based on CNN+Transformer by combining convolutional neural networks (CNN) and Transformer model. The model makes use of the translation and rotation invariance of CNN model to alleviate the variability of speech signals, and combines the multi head self attention mechanism of Transformer model to take into account the dependency between speech features and distance feature frames, thereby improving the parallel computing ability of the model, and realizing efficient and accurate automatic speech translation of English translators. Finally, the effectiveness of the proposed CNN+Transformer model is verified by simulation on the CCMT2019-BSTC dataset. The simulation results show that the proposed CNN+Transformer model has a good automatic voice translation effect. The BLEU value is 19.23, the parameter size and training duration are 18.99 M and 5.4 min respectively. Compared with the current commonly used end-to-end translation model seq2seq and cascading translation model NMT model and cascade model, the proposed CNN+Transformer model has better translation effect, and has certain effectiveness and superiority, which can be used for automatic voice translation of actual English translators.

关 键 词：端到端翻译模型 英语翻译器 自动翻译 CNN模型 Transformer模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]