Incorporating text information on presentation slides for spoken lecture retrieval

Abstract

In recent years, multimedia contents, such as audio and video data, have been increasing, and the search techniques for them are needed. Spoken content retrieval (SCR), which retrieves speech data using their transcription generated by an automatic speech recognition (ASR) system, has been widely studied. One of the challenges in SCR is how to deal with ASR errors found in the transcription. To alleviate the error effects, we propose to use the presentation slides along with the ASR results in the case of the retrieval targeting presentations taken place in an academic conference. We present three methods of using the presentation slides. We performed experiments on NTCIR-11 slide segmentation data set. In the results, we observed the significant improvement on the all proposed methods.