Abstract

Commercial interfaces for accessing digital speech data are based on ‘tape recorder ’ metaphors. However, such interfaces make it highly laborious to access complex speech data. The absence of effective interfaces is a major obstacle to exploiting the increasing number of speech archives now available online. More novel research interfaces provide potentially more effective access by presenting visual or textual indices into the underlying speech data. The current experimental study evaluates the utility of these newer techniques compared with a ‘tape recorder ’ interface. We compare: (a) High-level Visual Overviews showing the distribution and density of user query terms; (b) Textual Transcripts generated using state of the art ASR; (c) a tape recorder baseline. Laboratory tests showed that, contrary to our expectations, high-level visual information proved more useful than textual information, although both perform better than a tape-recorder baseline. Visual overviews enable users to quickly identify relevant regions to be played. In contrast, Textual transcripts can mislead users who try to extract detailed information solely by reading the transcript, without listening to the underlying speech.

Citations