By SVETLANA STENCHIKOVA

On the Saturday preceding Interspeech 2006, forty young researchers from around the globe came together for the second meeting of the Young Researchers Roundtable on Spoken Dialogs. The goal of the Roundtable was the search for the "holy grail" of dialog systems. The Round Table featured presentations and discussions on a range of topics including paralinguistic phenomena, evaluation of dialog systems, statistical methods, rapid development, and more.

Discussion of dialog system evaluation included spoken dialog system architectures ranging from simple VXML-based form filling to more sophisticated natural language understanding (as in TRIPS) to dialog management (as in Ravenclaw) technologies. Other areas in the speech/NLP community are benefiting from the yearly competitions which drive the researchers to create new methods and perfect them. An idea of dialog system competition was proposed but the obvious difficulties arise. Most significantly, how to compare

• systems representing different domains

• systems that are based on different architectures

 The technology used to build a system and the system's domain is tightly interdependent. A simple form filling system is more natural to be implemented using a VXML interface, while more complex decision-based system is not.

A question was raised "How can we make the dialog systems more popular?" Currently many people prefer speaking to a representative and try to skip system's menus to get connected directly to a live representative. We differentiated between the two types of systems based on the purpose:

• systems that replace the functionality that previously was done by humans (automated call centers)

• systems that add value for the user (e.g. CMU's Let's Go system).

The conclusion of this discussion was that the systems that benefit the caller by either saving time, giving an opportunity to access otherwise unavailable information are the types of systems would be preferred by users.

Many ideas about the Next Killer Application were also discussed:

• command and control dialogs for cell phones

• dialogs on multimodal devices

• "smart house" helping elderly.

• In-car control applications with wide range recognition features

• Video games where AI game characters engage in intelligent dialogs

• Developing world application, providing information access for the third-world countries

• Survey applications

• Dialogs for Robots

The consensus on the question of the "Next Killer Application" has not been reached, although everyone agreed that a breakthrough in ASR technology will give a new life to the research on dialog systems.