Mixing and merging for spoken document retrieval

This paper describes a number of experiments that explo- red the issues surrounding the retrieval of spoken documents. Two such issues were examined. First, attempting to find the best use of speech recogniser output to produce the highest retrieval effectiveness. Second, investigating the potential problems of retrieving from a so-called "mi- xed collection", i.e. one that contains documents from both a speech recognition system (producing many errors) and from hand transcription (producing presumably near perfect documents). The result of the first part of the work found that merging the transcripts of multiple recognisers showed most promise. The investigation in the second part showed how the term weighting scheme used in a retrieval system was important in determining whether the system was affected detrimentally when retrieving from a mixed collection.