Speech and Hand Transcribed Retrieval

Abstract

This paper describes the issues and preliminary work involved in the creation of an information retrieval system that will manage the retrieval from collections composed of both speech recognised and ordinary text documents. In previous work, it has been shown that because of recognition errors, ordinary documents are generally retrieved in preference to recognised ones. Means of correcting or eliminating the observed bias is the subject of this paper. Initial ideas and some preliminary results are presented. General TermsMeasurement, Experimentation.

Search of Spoken Documents Retrieves Well Recognized Transcripts

“…Although the trend is consistent across the data sets, it is not immediately clear what the cause of such a trend is: one explanation is that top ranked documents tend to contain a broader range of query words than those documents ranked lower. Another explanation mentioned by Sanderson and Shou (2002) is that transcripts of spoken documents containing query words assigned a high tf weight -which tend to be ranked highly by retrieval systems -often have a lower overall WER. Determining which of these possible causes might explain the observed effect was the subject of the next experiment.…”