Abstract: Excellent lip corpora can provide a good foundation for lip recognition. However, the lack of universal lip corpus is one of the important reasons for the slow development of lip recognition. This paper comprehensively reviews the characteristics of more than twenty corpora. Firstly, it briefly introduces the traditional methods and deep learning methods of lip recognition. Secondly, it focuses on the collection and research of the influential lip corpora for more than two decades, makes comparisons and analysis from twelve aspects, such as identification object, corpus size, recording mode and recording environment, and finds out their respective advantages, disadvantages and application scopes of various corpora, so that lip reading researchers will find it easier to choose corpora suitable for their own research direction. Then, it compares the algorithms adopted by various corpora and their performance. Finally, the difficulties faced by lip recognition are analyzed, and the future work is prospected.

Key words: lip recognition, lip corpus, deep learning, cross field

摘要： 优秀的语料库能为唇语识别提供良好的基础保障，但通用语料库的缺乏是导致唇语识别发展缓慢的重要原因之一。较为全面地综述了20多种语料库的相关特性。简单介绍了唇语识别的传统方法和深度学习方法。重点整理了近20多年较有影响力的唇语语料库，从识别对象、语料规模、录制方式与录制环境等12个方面进行比较分析，得出各种语料库的优缺点及适用范围，方便唇读工作者快速找到适合自己研究方向的语料库。比较了各种语料库采用何种算法及其所能达到的性能。对唇读面临的困难进行了剖析，对未来工作进行了展望。

关键词: 唇语识别, 唇语语料库, 深度学习, 交叉领域