Design and evaluation of a speech interface for remote database searching

Abstract

Advances in speech technology have made it possible to use speech as an effective man-machine interface in infor mation retneval applications. We believe that since it is among the most natural means of communication, limited use of speech should increase the speed and ease of access to large document databases. The purpose of this paper is to descnbe a two-year research programme which involved the design, im plementation and evaluation of a speech interface for the British Library Blaise online service The project began by examining an existing pnmitive voice interface which was developed during the early 1980s for an in-house text retneval system known as MicroBIRD. Evaluation of this interface has provided invaluable insights into the problems of voice interface design for online searching. The main lessons learned from this initial study were: (a) take full advantage of the well defined syntax of the query language to limit the difficulty of the speech recognition process; and (b) avoid antagonising the user by providing full control of the configuration of the interface, enabling varying degrees of audio reinforcement of visually presented data. Based on the experience gained from the MicroBiRD inter face we embarked on the more challenging task of designing a speech interface for the Blaise system. We will outline the hardware configuration, software development, analysis of the Blaise query language syntax and design features of the new interface. Having successfully developed the system the next logical step was to study the reactions of users to the interface. particularly in relation to the preferred mix of keyboard/voice input and screen/speech output. We have carried out a series of expenments using subjects from a wide range of back grounds. The evaluation expenments have shown that the use of voice for the input of commands and associated parameters is the area of greatest advantage of a voice interface. Indeed voice input is almost as fast as keyboard input, and sometimes slightly faster. Speech output, on the other hand, should be used pnmanly for providing short prompts (e.g. help messages) to the user. The reading out of an entire bibliographic record would be irksome and, because of the senal nature of speech. far too slow. If speech output is desired it should be limited to selected parts of a record such as author and title fields. Finally. our results have also shown that currently available speech recognition and synthesis hardware, along with intelli gent software, can provide an interface well suited to the needs of online information retrieval systems.