Investigation of efficient semi-automatic correction method using STD for automatic captioning
Yuji Terada, Tamiya Kenta, Atsuhiko Kai · 2017
Captioning lecture speech is very useful for better understanding. However, it takes high cost to do real-time manual captioning or even if we employ automatic speech recognition system and human correction together. In this paper, we propose a method to reduce a cost for human correction as a prerequisite of a framework for captioning using automatic speech recognition system. Specifically, we investigate the effect of incorporating a simple human's feedback which only includes error words such as technical terms and proper nouns, and identifying and correcting the caption text by using spoken term detection system. Moreover, we investigate the method to improve the accuracy of the automatic captioning system by using the corrected text for semi-supervised language model adaptation. Throughout the preliminary experiments, it was found that the proposed caption correcting system could improve the word error rate.