Improvement of Spoken Term Detection Technique and its Application to Speech Recognition and Spoken Document Retrieval

点击翻译按钮获取中文摘要

The goal of this study is refinement of the spoken term detection (STD) technique. An STD technology can detect speech intervals, where query terms are uttered, in lots of spoken documents. In addition, I also aim at adapting the STD method to other technologies such as speech recognition for improving their refinement. I developed the STD methods using multiple speech recognizers' outputs, confidence measures based on majority voting, and machine learning. In the experiment on STD, my techniques achieved improvement of the STD performance comparing to the baseline system. In addition, I adopted my STD technique to making recognition dictionary, which is necessary for speech recognition. In the result, my technique improves speech recognition performance. Furthermore, I implemented the STD technique to an electronic note-taking support system and evaluated its effectiveness of utilization of the STD technique. The system with STD is useful for retrieving words from multimedia data.