Lyrebird™: Developing Spoken Dialog Systems Using Examples

Abstract

An early release software product for the rapid development of spoken dialog systems (SDS's), known as Lyrebird TM [1][2] [3], will be demonstrated that makes use of grammatical inference to build natural language, mixed initiative, speech recognition applications.The demonstration will consist of the presenter developing a spoken dialog system using Lyrebird TM , and will include a demonstration of some features that are still in the prototype phase.Using Lyrebird TM , developers build a spoken dialog system with the following steps: 1. Firstly the dialog is specified. This is typically developed using a drag and drop GUI that defines a computer directed menu driven application (See Figure. 1). Alternatively a wizard can be used whereby the developer defines the application in terms of slots that need to be filled. 2. The dialog description is then generalised to include mixed initiative input.Starting grammars are automatically created that cover computer directed input. 3. The developer is then prompted for text describing selected phrases that a speaker would use when interacting with the developed SDS (See Figure 2.) Mixed-initiative natural-language grammars suitable for use with a speechrecognition system are then inferred, using grammatical inference.If desired, the developer can subsequently improve the system under development, using a range of interfaces. For instance:1. The developer can teach Lyrebird how the speaker will behave using a tool known as the simulator. The simulator enables the user to interact with the application under development using text. (See Figure 3.) Out of grammar phrases are tagged using either robust parsing, or manual input from the developer. 2. The developer can teach Lyrebird how the system should behave using a tool known as the scenario editor (See Figure 4.) Using the scenario editor sample interactions are presented to the developer in a text format similar to that used

Learning tree languages from text

“…In the future, we will try to compare our work with other works on the inference of tree languages and of context-free languages, as they are contained, e.g., in [5,27,30,33,43,47,48,61,[64][65][66][71][72][73]. Moreover, it would be interesting to extend the work to other, more general classes of tree languages and the corresponding languages of yielded strings, see [70] for a short exposition.…”

A Case Study of Applying Decision Theory in the Real World

“…In addition, numerous other aspects of the spoken dialog system can be "learned" for a specific task, such as application-specific grammars [64,65], prompt wording [16,45], choice of text-to-speech audio [11], and others. Learning in these areas can certainly improve performance of a spoken dialog system, but is separate from the dialog management task.…”