Although many recent dialog systems have shown that we can achieve effective spoken interaction with a computer, they tend to target the "average" portions of the population, those whose speech and hearing fall within the norm of the whole population. This has taught us much about how the dialog must change when people interact with a computer instead of a human. We have developed system architectures capable of finding and presenting useful information for the average user, but these systems cannot be used by everyone. People who are, for some reason, considered to be outliers in the general population cannot yet access the information these systems provide. The objective of the Let's Go project is to create a basic dialog system that we can use to test how to extend system access to extreme populations. The portions of the population that we have chosen as representatives of the extreme are the elderly and non-native speakers of English.
As we age, perception is lessened, attention is narrowed and memory is limited. This makes it extremely difficult to listen to and use the information a dialog system furnishes. When we speak a foreign language, we often have not mastered all of its sounds or its grammatical constructions. This makes it difficult for a dialog system to understand what we want. Our populations therefore complement each other since the elderly provide an extreme in the use of speech output and non-natives do the same for speech input.
Our specific interest in creating a dialog system for such populations came first from the observation that elderly visitors apparently had a much harder time using our spoken dialog systems than younger users. To investigate this, we devised a simple experiment with elderly visitors to CMU's Homecoming testing their comprehension of natural and synthetic speech over the telephone under a number of conditions. The results [1] show a drop in comprehension as age increases.
In this paper, we will describe the basic dialog system that we have created to serve as our testbed. Let's Go has its roots in the CMU Communicator [2] system architecture. From that starting point and with our experimental goals in mind, we have made modifications to the basic architecture, making it easier to change necessary parts of the system as we adapt to the new populations, such as making the parser more tolerant to grammatical errors. With the system now in place we are experimenting with ways to: enhance the speech output so that the elderly can understand it better; detect what a non-native speaker meant to say and offer hints of how to say it better.
Our system provides bus schedule information for the city of Pittsburgh. We are working with the Port Authority Transit System (PAT) to use their bus schedules and recordings of actual calls to their help desk to build our system.