论文部分内容阅读
人工智能(AI)正在入侵唇读领域。Google的DeepMind和牛津大学的一个合作项目将深度学习应用到BBC的一个庞大的数据集中,以创建一个唇部阅读系统。唇读是人类一项独特的技艺,也是非常困难的一件事,它对于语言语境和知识理解的要求并不亚于视觉上的线索,然而AI又做到了。该AI系统从6个不同的电视节目,包括Newsnight,BBC Breakfast和Question Time的约5 000小时的节目中进行训练。这些视频总共包含118 000个句子。牛津大学和Deep Mind研究人员先是在2010年1月至2015年12月期间播出的节目上对
Artificial Intelligence (AI) is invading the field of lip reading. A partnership between Google’s DeepMind and the University of Oxford will apply deep learning to a massive data set of the BBC to create a lip reading system. Lip reading is a unique human skill, but also a very difficult one. Its requirement of linguistic context and knowledge comprehension is no less than visual clues, however AI does it again. The AI system trains from about 6,000 different shows, including Newsnight, BBC Breakfast and Question Time, for about 5,000 hours. The videos contain a total of 118,000 sentences. Oxford University and Deep Mind researchers first broadcast programs from January 2010 to December 2015