Some aspects of evaluating the performance of a speech recognition system in real applications

The application of small speech recognition machines has proved useful in some domains of voice input for command and control. But it seems to be rather difficult to find adequate methods for evaluating and comparing the performance of such systems. By means of a formatted data entry job conducted with the DAWID speech recognition system we investigated mutual dependencies of some user-specific and system-specific factors and their influence on the efficiency of the total system. In this context we studied the time behaviour of the system in more detail.