基于Transformer的越南语连续语音识别  

Vietnamese Continuous Speech Recognition Based on Transformer

作　　者：刘佳文 屈丹[1] 杨绪魁 张昊[1] 唐君 LIU Jiawen;QU Dan;YANG Xukui;ZHANG Hao;TANG Jun(Information Engineering University, Zhengzhou 450001, China)

机构地区：[1]信息工程大学,河南郑州450001

出　　处：《信息工程大学学报》2020年第2期129-133,152,共6页Journal of Information Engineering University

基　　金：国家自然科学基金资助项目(61673395)。

摘　　要：针对现有越南语语音识别模型大量使用循环神经网络,无法并行训练,模型收敛速度慢的问题,构建基于transformer的端到端语音识别模型,加快模型训练的同时,减少对语言学知识的依赖。通过分析越南语发音规律,系统以音素作为识别单元,对识别结果利用基于音节的语言模型进行重打分。实验结果表明,基于transformer的越南语语音识别系统与其他方法相比,能够同时提升识别率和收敛速度。To solve the problem that the existing Vietnamese speech recognition model has slow convergence and uses a large number of recurrent neural networks which cannot be trained in parallel,this paper builds an end-to-end speech recognition model based on transformer.The model accelerates the training and reduces dependence on linguistic knowledge.By analyzing the pronunciation rules of Vietnamese,phonemes are used as recognition units,and a syllable-based language model is constructed for rescoring.Experimental results show that the Vietnamese speech recognition system based on the transformer proposed in this paper can improve the recognition rate and convergence speed at the same time compared with other methods.

关 键 词：越南语 TRANSFORMER 端到端 语音识别 

分 类 号：TN912.34[电子电信—通信与信息系统]