Time-compressing speech: ASR transcripts are an effective way to support gist extraction

Abstract

A major problem for users exploiting speech archives is the laborious nature of speech access. Prior work has developed methods that allow users to efficiently identify and access the gist of an archive using textual transcripts of the conversational recording. Text processing techniques are applied to these transcripts to identify unimportant parts of the recording and to excise these, reducing the time taken to identify the main points of the recording. However our prior work has relied on human-generated as opposed to automatically generated transcripts. Our study compares excision methods applied to human-generated and automatically generated transcripts with state of the art word error rates (38%). We show that both excision techniques provide equivalent support for gist extraction. Furthermore, both techniques perform better than the standard speedup techniques used in current applications. This suggests that excision is a viable technique for gist extraction in many practical situations. © 2008 Springer-Verlag Berlin Heidelberg.