低资源少样本连续语音识别最新进展  

Overview of Recent Progress in Low-resource Few-shot Continuous Speech Recognition

作　　者：屈丹[1] 杨绪魁 闫红刚[1] 陈雅淇 牛铜[1] QU Dan;YANG Xukui;YAN Honggang;CHEN Yaqi;NIU Tong(School of Information System Engineering,Strategic Support Force Information Engineering University,Zhengzhou 450001,China)

机构地区：[1]战略支援部队信息工程大学信息系统工程学院,河南郑州450001

出　　处：《郑州大学学报（工学版）》2023年第4期1-9,共9页Journal of Zhengzhou University（Engineering Science）

基　　金：国家自然科学基金资助项目(62171470);河南省中原科技创新领军人才项目(234200510019)。

摘　　要：低资源少样本语音识别是目前语音识别行业面临的迫切技术需求。首先,总结了低资源连续语音识别技术的框架技术,重点介绍了低资源语音在特征提取、声学建模和资源扩展等方面的若干关键技术研究进展。其次,在连续语音识别框架技术发展的基础上,重点阐述了生成对抗网络、自监督表示学习、深度强化学习和元学习等高级深度学习技术在解决少样本语音识别方面的最新发展,如FGSM、wav2vec、AMS等代表性方法。在此基础上,分析了目前该技术面临的互补有限、数据和任务不均衡与模型轻量化部署问题。最后,对低资源少样本连续语音识别进行了总结,提出未来少样本训练识别的研究方向可以朝着先验信息引入、假设空间约束条件设定等方向进一步研究。Low-resource few-shot speech recognition is an urgent technical demand faced by the speech recognition industry.The framework technology for few-shot speech recognition was first briefly discussed in this study.The research progress of several important low resource speech technologies,including feature extraction,acoustic model,and resource expansion,was then highlighted.The latest advancements in deep learning technologies,such as generative adversarial networks,self-supervised representation learning,deep reinforcement learning,and meta-learning,were then focused on how to address few-shot speech recognition on the basis of the development of continuous speech recognition framework technology.On that basis,the problems of limited complementarity,unbalanced task and model deployment faced by this technology were analyzed for the subsequent development.Finally,a summary and prospects of few-shot continuous speech recognition were given.

关 键 词：低资源少样本 连续语音识别 生成对抗网络 自监督表示学习 深度强化学习 元学习 

分 类 号：TN912.34[电子电信—通信与信息系统]