There are several types of survey work in traffic engineering and transport where it is highly desirable to record field data by speech. This may be because of the speed of events, the enumerator's need to keep his eyes on the phenomenon of interest and/or his hands free for the operation of equipment, or the rigours of a hostile environment. Where this involves recording onto audio tape, the resulting transcription task can be daunting and inefficient. Automatic speech recognition equipment offers a potential solution to this problem. Previous research at Leeds University showed that although high levels of transcription accuracy could be achieved in laboratory and field trials, performance was unpredictable. Therefore considering the complexity of the operational parameters and the high cost of equipment, it was concluded that speech recognition was not suitable for the applications tested at that time. Now that more advanced speech recognisers are available and the purchase price is considerably lower, a further study has been undertaken to reevaluate the use of speech recognition equipment. In this paper the problems of the speech interface and the design of appropriate experiments are discussed. The initial results of tests applied to vehicle registration plate data using an automatic speech recogniser are also presented. Early indications are that accuracy levels are still insufficient and costs still too high.
Abstract