基于链接时序分类的日语语音识别  

Towards Connectionist Temporal Classification Speech Recognition System for Japanese

作　　者：孙健[1] 郭武[1] SUN Jian;GUO Wu(University of Science and Technology of China,National Engineering Laboratory for Speech and Language Information Processing,Hefei 230027,China)

机构地区：[1]中国科学技术大学语音及语言信息处理国家工程实验室,合肥230027

出　　处：《小型微型计算机系统》2018年第10期2129-2133,共5页Journal of Chinese Computer Systems

基　　金：国家重点研发计划专项项目(2016YFB1001303)资助

摘　　要：目前,端到端的语音识别系统因其简洁性和高效性成为大规模连续语音识别的发展趋势.本文将基于链接时序分类的端到端技术应用到日语语音识别上,考虑到日语中平假名、片假名和日语汉字多种书写形式的特性,通过在日语数据集上的实验,探讨了不同建模单元对识别性能的影响;进一步将音素信息应用到模型的初始网络训练中,改善语音识别系统性能,最终效果优于基于隐马尔可夫模型和双向长短时记忆网络的主流语音识别系统.The end-to-end framework has become the state-of-the-art method in large vocabulary continuous speech recognition （ LVC- SR ） because of its simplicity and efficiency. In this paper, the end-to-end technology based on Connectionist Temporal Classification （ CTC ）is applied to Japanese speech recognition. Considering the characteristic of various written forms among hiragana, katakana and kanji in Japanese, we discuss the impact of different modeling units on recognition performance through experiments on Japanese dataset. Then we combine phoneme information into the acoustic model to improve the performance. Experiments demonstrate the ef- fectiveness of the proposed methods, which can achieve better performance than the mainstream speech recognition system based on Hidden Markov Model and Bi-directional long-short memory network.

关 键 词：语音识别 日语 链接时序分类 端到端 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]