Design of Conversion Module from Korean Video to English Text for English Subtitle Generator Using Hidden Markov Model Algorithm

Abstract

Based on Youtube data on 2011, total views of Hallyu or Korean cultures related videos have reached 2.3 billion views, thrice compared to 2010 (800 million views). The rising popularity of Korean culture takes effect on the emergence of new English subtitle forums for Korean videos. These forums customers have to wait for minimum 2 days to get new video subtitle. In order to solve this issue, this research is held to build an automatic English subtitle generator for Korean video by combining and implementing today technologies called SubGen (the abbreviation of Subtitle Generator). By using SubGen, customers don’t have to wait days to get English subtitle for Korean videos. We hope this can give contribution for Korean natural language processing technology field. SubGen receives video as input and generate Subrip Text subtitle file (.srt) as output. SubRip Text file contains timing and text spoken for video subtitle. This research is focused on the Korean speech to text process by using the implementation of semi-continuous Hidden Markov Model (SCHMM) algorithm. Training process for this model use phoneme segmentation and forward-backward algorithm in HMM parameter estimation. Resources data for training and transcription are Korean pronunciation dictionary in Sphinx format using hangul syllable, 25 words including numeric and places in 225 sample audio and transcription texts. The training results in 36% accuracy for SCHMM implementation. The overall phases use 16 Korean video varied in content, speakers, and duration as first input. The test results for video to audio conversion, Korean to English translation and subtitle file creation tests is succeed. While speech to text process has 31% accuracy.