摘要
利用云端语音识别引擎和机器翻译引擎,结合开源语音处理软件ffmpeg,设计并实现了一个科普视频汉英双语字幕生成的系统。将科普视频文件用开源软件提取音频内容,调用百度云端语音识别引擎(https://aip.baidubce.com/)联合汉语科普知识库,实现语音到汉语字幕及其时间线的转换;调用百度云端机器翻译引擎(http://api.fanyi.baidu.com/)联合汉英科普对译库,将汉语字幕翻译为英文字幕,并对应到汉语字幕的时间线上,最后生成科普视频的汉英双语云端语音识别字幕。本文利用真实科普视频评估了本文所提系统的处理能力,从汉语语音到英文字幕总正确(可懂)率为77.3%;进一步分析该字幕生成系统的人工用时,接近全人工处理的1/5,能够有效降低人工成本,提高科普视频汉英双语字幕的生成效率。
In this paper,a Chinese English subtitle generation system for popular science videos is designed and implemented,using engines in cloud for automatic speech recognition and machine translation and combining with open source speech processing software ffmpeg.We extracted audio content from popular science video files with open source software,used the cloud engines(https://aip.baidubce.com/)and Chinese popular science knowledge base to get Chinese subtitles and their timelines,and formed final Chinese subtitles.Then,we translated Chinese subtitles into English subtitles with the cloud engines(http://api.fanyi.baidu.com/)and Chinese English popular science translation data base,which were mapped to the timeline of Chinese subtitles.And we generated Chinese and English subtitles of popular science videos.We also evaluated the processing ability of the system by real popular science videos.The overall sentence accuracy(intelligibility)from Chinese voice to English subtitles is 77.3%.Further analysis showed that the manual time of the system is near to 1/5 of the total manual processing time,which can effectively reduce the labor cost and improve the efficiency of generating Chinese English subtitles for popular science videos.
作者
周城光
周军
韦向峰
周文佳
王荣泉
ZHOU Chengguang;ZHOU Jun;WEI Xiangfeng;ZHOU Wenjia;WANG Rongquan(Language and Intelligent Information Processing Laboratory,Institute of Acoustics,Chinese Academy of Sciences,Beijing,100190,China)
出处
《网络新媒体技术》
2023年第2期62-68,共7页
Network New Media Technology
关键词
科普
视频字幕
语音识别
机器翻译
popularization of science
video subtitles
speech recognition
machine translation
