Wildfire Communications: "What can I do for you?"

RELease 1.0, Oct 19, 1994

For example, at any point in a dialog there are only a few things the user is expected to say, usually six and, if possible, not more than three dozen words. The developers phrase Wildfire's questions in a way that elicits a "yes" or a "no," to avoids mixups with "yup," "sure" and "nope." Wildfire also treats many phrases as single utterances, which improves accuracy. A phrase such as "What's it say?" and a person's full name are stored and treated as single words. At any time, a user can say "Never mind" or "What are my options?" Because Wildfire usually has a good idea of what to expect from a caller, she can turn on one or more specialized recognizers (from Voice Processing Corporation of Cambridge, MA). A speaker-independent, continuous digit recognizer deals with phone numbers; a speaker-independent discrete recognizer handles common utterances, such as most system commands; and a speaker-dependent, discrete recognizer handles utterances such as contact and object names. Wildfire is developing a speaker-independent, continuous speech recognizer specifically for dates and times.

Although this recognition process requires a lot of ping-ponging between caller and agent, the interaction still seems smooth and natural.

Software architecture and the Virtual Hallway

The Wildfire system uses Object Design's ObjectStore database. The system's principal constructs (objects) are "assistants," "sessions" and "gadgets." Assistants and sessions we just described. Gadgets are abstractions for devices, including telephones, pagers, fax machines, e-mail, PDAs and computers. In its first release, Wildfire supports only phones and pagers. Fax capabilities are easy to add, and the system design will support the rest, but the other devices and media will require more complex integration. For example, Wildfire's contact list can hold information on 150 people. Wildfire developers want to make it easy to import names and numbers from your favorite PIM to speed setup.

The system's designers have built many layers of abstraction into its multimedia architecture. Although Wildfire's initial interface is primarily speech, all objects have a canonical form (Wildfire's engineers call them "memes") with alternate values underneath; the system chooses the appropriate one given the user's preferences and session capabilities. That means a prompt or response can be typed, spoken or some combination. It also makes it easy to do different languages or provide more verbose responses to novice users.

Every command has a touch-tone equivalent, which not only means that users can issue commands from a phone keypad if they need to be discreet, or if there's too much background noise, but also that any device that can generate touch tones can drive Wildfire. Eventually, different kinds of computer hardware may play a larger role in the Wildfire system. PDAs and PCs could display options and enhance users' interactions with the system. However, because users can do everything through speech from any phone, devices (even touch tone) are unnecessary.