Speech-recognition systems should be deployed in three progressive phases.
In the pilot deployment phase, only a few hundred calls are allowed into the system. The designers and programmers then tune up everything as needed ”the recognizer, the design of the user interface, any technical glitches, and perhaps even the prompts. The relatively low volume of calls can be analyzed manually. It's a little time-consuming , but easy to do and yields great payback.
Only move on to the partial deployment phase once you're certain that all known issues have been worked out and the system is running smoothly. The system takes more calls ”perhaps as many as 10,000 ”to get a larger sample of data over a longer period of time. Because of the larger volume, most of the calls are analyzed using automated methods , but you should still do some manual analysis of problem calls.
In the full deployment phase, the call volume is increased to the system's full capacity while the designers and programmers continue to monitor the system, albeit at longer intervals than before, using almost entirely automated means.
